Cheng-Han Chiang, Xiaofei Wang, Linjie Li, Chung-Ching Lin, Kevin Lin, Shujie Liu, Zhendong Wang, Zhengyuan Yang, Hung-yi Lee, Lijuan Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs) are increasingly used not only to generate code, but also to judge it: comparing, ranking, or scoring competing solutions.However, their reliability in this evaluative role remains poorly understood.Inconsistent or flawed judgments can undermine benchmarks and distort training signals.This paper investigates the performance and robustness of LLMs when used as code judges.We introduce CodeJudgeBench, a benchmark explicitly designed to evaluate LLM-as-a-Judge models across three critical coding tasks: code generation, code repair, and unit test generation.We comprehensively benchmark the performance of 26 LLM-as-a-Judge models, encompassing general-purpose, code-tuned, and reasoning models.Our empirical findings reveal that relatively small reasoning models (e.g., Qwen3-8B) can outperform much larger non-reasoning models up to 70B.We further stress-test robustness by applying both general and code-specific perturbations.All models show significant instability and are sensitive to changes such as response ordering, variable naming, and misleading comments.These findings highlight serious concerns about the consistency and robustness of LLM-based judges for coding tasks.

Tzu-Quan Lin, Heng-Cheng Kuo, Tzu-Chieh Wei, Hsi-Chun Cheng, Chun Wei Chen, Hsien-Fu Hsiao, Yu Tsao, Hung-yi Lee. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The LLM-as-a-judge paradigm uses large language models (LLMs) for automated text evaluation, which assigns a score to the text based on some scoring rubrics.Existing methods for LLM-as-a-judge use cross-entropy (CE) loss for fine-tuning, which neglects the numeric nature of score prediction.Recent work addresses numerical prediction limitations of LLM fine-tuning through regression-aware finetuning, which, however, does not consider chain-of-thought (CoT) reasoning for score prediction.In this paper, we introduce TRACT (Two-stage Regression-Aware fine-tuning with CoT), a method combining CoT reasoning with regression-aware training.The training objective of TRACT combines the CE loss for learning the CoT reasoning and the regression-aware loss for the score prediction.TRACT consists of two stages: first, a seed LLM is fine-tuned to generate CoTs; next, we retrain the seed LLM using the CoTs generated by the LLM trained in stage 1.Experiments across four LLM-as-a-judge datasets and two LLMs show that TRACT significantly outperforms existing methods.Extensive ablation studies validate the importance of each component in TRACT. 1

Guan-Ting Lin, Prashanth Gurunath Shivakumar, Aditya Gourav, Yile Gu, Ankur Gandhe, Hung-yi Lee, Ivan Bulyko. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We propose a self-refining framework that enhances ASR performance with only unlabeled datasets. The process starts with an existing ASR model generating pseudo-labels on unannotated speech, which are then used to train a highfidelity text-to-speech (TTS) system. Then, synthesized speech text pairs are bootstrapped into the original ASR system, completing the closed-loop self-improvement cycle. We demonstrated the effectiveness of the framework on Taiwanese-Mandarin speech. Leveraging 6,000 hours of unlabeled speech, a moderate amount of text data, and synthetic content from the AI models, we adapt Whisper-large-v2 into a specialized model, Twister. Twister reduces error rates by up to $20 \%$ on Mandarin and $50 \%$ on Mandarin-English code-switching benchmarks compared to Whisper. Results highlight the framework as a compelling alternative to pseudo-labeling self-distillation approaches and provides a practical pathway for improving ASR performance in lowresource or domain-specific settings.

Creating a unified speech and music model requires expensive pre-training. Model merging can instead create a unified audio model with minimal computational expense. However, direct merging is challenging when the models are not aligned in the weight space. Motivated by Git Re-Basin, we introduce a correlation-permutation approach that aligns a music encoder’s internal layers with a speech encoder. We extend previous work to the case of merging transformer layers. The method computes a permutation matrix that maximizes the model’s features-wise cross-correlations layer by layer, enabling effective fusion of these otherwise disjoint models. The merged model retains speech capabilities through this method while significantly enhancing music performance, achieving an improvement of 14.83 points in average score compared to the linear interpolation model merging. This work allows the creation of unified audio models from independently trained encoders.

Evaluation of text-to-music systems is constrained by the cost and availability of collecting experts for assessment. AudioMOS 2025 Challenge track 1 is created to automatically predict music impression (MI) as well as text alignment (TA) between the prompt and the generated musical piece. This paper reports our winning system, which uses a dual-branch architecture with pre-trained MuQ and RoBERTa models as audio and text encoders. A cross-attention mechanism fuses the audio and text representations. For training, we reframe the MI and TA prediction as a classification task. To incorporate the ordinal nature of MOS scores, one-hot labels are converted to a soft distribution using a Gaussian kernel. On the official test set, a single model trained with this method achieves a system-level Spearman’s Rank Correlation Coefficient (SRCC) of 0.991 for MI and 0.952 for TA, corresponding to a relative improvement of $21.21 \%$ in MI SRCC and $31.47 \%$ in TA SRCC over the challenge baseline.

Automatic Speech Recognition (ASR) models demonstrate outstanding performance on high-resource languages but face significant challenges when applied to low-resource languages due to limited training data and insufficient cross-lingual generalization. Existing adaptation strategies, such as shallow fusion, data augmentation, and direct fine-tuning, either rely on external resources, suffer computational inefficiencies, or fail in test-time adaptation scenarios. To address these limitations, we introduce Speech Meta In-Context LEarning (SMILE), an innovative framework that combines meta-learning with speech in-context learning (SICL). SMILE leverages metatraining from high-resource languages to enable robust, few-shot generalization to low-resource languages without explicit fine-tuning on the target domain. Extensive experiments on the ML-SUPERB benchmark show that SMILE consistently outperforms baseline methods, significantly reducing character and word error rates in training-free few-shot multilingual ASR tasks.

Large Audio-Language Models (LALMs) can take audio and text as the inputs and answer questions about the audio. While prior LALMs have shown strong performance on standard benchmarks, there has been alarming evidence that LALMs can hallucinate what is presented in the audio. To mitigate the hallucination of LALMs, we introduce Audio-Aware Decoding (AAD), a lightweight inference-time strategy that uses contrastive decoding to compare the token prediction logits with and without the audio context. By contrastive decoding, AAD promotes the tokens whose probability increases when the audio is present. We conduct our experiment on object hallucination datasets with three LALMs and show that AAD improves the F1 score by 0.046 to 0.428. We also show that AAD can improve the accuracy on general audio QA datasets like Clotho-AQA by 5.4% to $\mathbf{1 0. 3 \%}$. We conduct thorough ablation studies to understand the effectiveness of each component in AAD.

Despite progress in end-to-end ASR, real-world domain mismatches still cause performance drops, which Test-Time Adaptation (TTA) aims to mitigate by adjusting models during inference. Recent work explores combining TTA with external language models, using techniques like beam search rescoring or generative error correction. In this work, we identify a previously overlooked challenge: TTA can interfere with language model rescoring, revealing the nontrivial nature of effectively combining the two methods. Based on this insight, we propose SUTA-LM, a simple yet effective extension of SUTA, an entropy-minimization-based TTA approach, with language model rescoring. SUTALM first applies a controlled adaptation process guided by an auto-step selection mechanism leveraging both acoustic and linguistic information, followed by language model rescoring to refine the outputs. Experiments on 18 diverse ASR datasets show that SUTA-LM achieves robust results across a wide range of domains.11The source code is available at https://github.com/hhhaaahhhaa/ASR-TTA

Accurate audio quality estimation is essential for developing and evaluating audio generation, retrieval, and enhancement systems. Existing non-intrusive assessment models predict a single Mean Opinion Score (MOS) for speech, merging diverse perceptual factors and failing to generalize beyond speech. We propose MMMOS, a no-reference, multidomain audio quality assessment system that estimates four orthogonal axes: Production Quality, Production Complexity, Content Enjoyment, and Content Usefulness across speech, music, and environmental sounds. MMMOS fuses frame-level embeddings from three pretrained encoders (WavLM, MuQ, and M2D) and evaluates three aggregation strategies with four loss functions. By ensembling the top eight models, MMMOS shows a $\mathbf{2 0 - 3 0 \%}$ reduction in mean squared error and a 4-5% increase in Kendall’s $\tau$ versus baseline, gains first place in six of eight Production Complexity metrics, and ranks among the top three on 17 of 32 challenge metrics.

Speech emotion recognition (SER) systems often exhibit gender bias. However, the effectiveness and robustness of existing debiasing methods in such multi-label scenarios remain underexplored. To address this gap, we present EMO-Debias—a large-scale comparison of 13 debiasing methods applied to multi-label SER. Our study encompasses techniques from pre-processing, regularization, adversarial learning, biased learners, and distributionally robust optimization. Experiments conducted on acted and naturalistic emotion datasets, using WavLM and XLSR representations, evaluate each method under conditions of gender imbalance. Our analysis quantifies the trade-offs between fairness and accuracy, identifying which approaches consistently reduce gender performance gaps without compromising overall model performance. The findings provide actionable insights for selecting effective debiasing strategies and highlight the impact of dataset distributions.

Recent attempts at source tracing for codecbased deepfake speech (CodecFake), generated by neural audio codec-based speech generation (CoSG) models, have exhibited suboptimal performance. However, how to train source tracing models using simulated CoSG data while maintaining strong performance on real CoSG-generated audio remains an open challenge. In this paper, we show that models trained solely on codec-resynthesized data tend to overfit to non-speech regions and struggle to generalize to unseen content. To mitigate these challenges, we introduce the Semantic-Acoustic Source Tracing Network (SASTNet), which jointly leverages Whisper for semantic feature encoding and Wav2vec2 with AudioMAE for acoustic feature encoding. Our proposed SASTNet achieves state-of-theart performance on the CoSG test set of CodecFake+ dataset, demonstrating its effectiveness for reliable source tracing.

Spoken dialogue modeling poses challenges beyond text-based language modeling, requiring real-time interaction, turn-taking, and backchanneling. While most Spoken Dialogue Models (SDMs) operate in half-duplex mode—processing one turn at a time—emerging full-duplex SDMs can listen and speak simultaneously, enabling more natural conversations. However, current evaluations remain limited, focusing mainly on turn-based metrics or coarse corpus-level analyses. To address this, we introduce Full-Duplex-Bench, a benchmark that systematically evaluates key interactive behaviors: pause handling, backchanneling, turn-taking, and interruption management. Our framework uses automatic metrics for consistent, reproducible assessment and provides a fair, fast evaluation setup. By releasing our benchmark and code, we aim to advance spoken dialogue modeling and foster the development of more natural and engaging SDMs.

Transformer-based self-supervised models have achieved remarkable success in speech processing, but their large size and high inference cost present significant challenges for real-world deployment. While numerous compression techniques have been proposed, inconsistent evaluation metrics make it difficult to compare their practical effectiveness. In this work, we conduct a comprehensive study of four common compression methods, including weight pruning, head pruning, low-rank approximation, and knowledge distillation on self-supervised speech Transformers. We evaluate each method under three key metrics: parameter count, multiply-accumulate operations, and real-time factor. Results show that each method offers distinct advantages. In addition, we contextualize recent compression techniques, comparing DistilHuBERT, FitHuBERT, LightHuBERT, ARMHuBERT, and STaRHuBERT under the same framework, offering practical guidance on compression for deployment.

Modern speech quality prediction models are trained on audio data resampled to a specific sampling rate. When tested on audio with a higher sampling rate, these models can produce biased scores. We present HighRateMOS, the first non-intrusive mean opinion score (MOS) model that explicitly considers sampling rate. HighRateMOS ensembles three model variants that exploit the following information: (i) a learnable embedding of speech sampling rate, (ii) Wav2vec 2.0 selfsupervised embeddings, (iii) multi-scale CNN spectral features, and (iv) MFCC features. In AudioMOS 2025 Track 3, HighRateMOS ranked first in five of eight metrics. Our experiments confirm that modeling sampling rate leads to more robust and sampling-rate-agnostic speech quality predictions.

Bias in speech emotion recognition (SER) systems often stems from spurious correlations between speaker characteristics and emotional labels, leading to unfair predictions across demographic groups. Many existing debiasing methods require model-specific changes or demographic annotations, limiting their practical use. We present COVADA, a Confidence-Oriented Voice Augmentation Debiasing Approach that mitigates bias without modifying model architecture or relying on demographic information. CO-VADA identifies training samples that reflect bias patterns present in the training data and then applies voice conversion to alter irrelevant attributes and generate samples. These augmented samples introduce speaker variations that differ from dominant patterns in the data, guiding the model to focus more on emotion-relevant features. Our framework is compatible with various SER models and voice conversion tools, making it a scalable and practical solution for improving fairness in SER systems.

Real-world audio often mixes speech and music, yet models typically handle only one domain. This paper introduces a multi-teacher distillation framework that unifies speech and music models into a single one while significantly reducing model size. Our approach leverages the strengths of domain-specific teacher models, such as HuBERT for speech and MERT for music, and explores various strategies to balance both domains. Experiments across diverse tasks demonstrate that our model matches the performance of domain-specific models, showing the effectiveness of cross-domain distillation. Additionally, we conduct few-shot learning experiments, highlighting the need for general models in real-world scenarios where labeled data is limited. Our results show that our model not only performs on par with specialized models but also outperforms them in few-shot scenarios, proving that a cross-domain approach is essential and effective for diverse tasks with limited data. Code and models are released at https://github.com/johnwei0325/Multi-Distillation-from-Speech-and-Music-Representation-Models

Understanding the internal mechanisms of large audio-language models (LALMs) is crucial for interpreting their behavior and improving performance. This work presents the first in-depth analysis of how LALMs internally perceive and recognize auditory attributes. By applying vocabulary projection on three state-of-the-art LALMs, we track how attribute information evolves across layers and token positions. We find that attribute information generally decreases with layer depth when recognition fails, and that resolving attributes at earlier layers correlates with better accuracy. Moreover, LALMs heavily rely on querying auditory inputs for predicting attributes instead of aggregating necessary information in hidden states at attribute-mentioning positions. Based on our findings, we demonstrate a method to enhance LALMs. Our results offer insights into auditory attribute processing, paving the way for future improvements.

Yi-Cheng Lin, Kang-Chieh Chen, Zhe-Yan Li, Tzu-Heng Wu, Tzu-Hsuan Wu, Kuan-Yu Chen, Hung-yi Lee, Yun-Nung Chen. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

With advancements in large audio-language models (LALMs), which enhance large language models (LLMs) with auditory capabilities, these models are expected to demonstrate universal proficiency across various auditory tasks.While numerous benchmarks have emerged to assess LALMs' performance, they remain fragmented and lack a structured taxonomy.To bridge this gap, we conduct a comprehensive survey and propose a systematic taxonomy for LALM evaluations, categorizing them into four dimensions based on their objectives: (1) General Auditory Awareness and Processing, (2) Knowledge and Reasoning, (3) Dialogue-oriented Ability, and (4) Fairness, Safety, and Trustworthiness.We provide detailed overviews within each category and highlight challenges in this field, offering insights into promising future directions.To the best of our knowledge, this is the first survey specifically focused on the evaluations of LALMs, providing clear guidelines for the community.We will release the collection of the surveyed papers and actively maintain it to support ongoing advancements in the field.

Multilingual Automatic Speech Recognition (ASR) aims to recognize and transcribe speech from multiple languages within a single system. By leveraging a vast amount of data and incorporating language tokens as prefixes to guide the recognition process, Whisper is one of the most advanced multilingual ASR models. However, despite its success, Whisper struggles with unseen languages, which are not included in its pre-training. Motivated by the observation that many languages share linguistic characteristics, we propose a method that exploits these relationships to improve ASR performance of Whisper in unseen languages. With Whisper’s predicted language probabilities as weight, the approach weighted sums the embeddings of language tokens . Based on this method, we also develop a predictor-based approach that supports our assumption about the weighted sum method. Our proposed methods demonstrate substantial improvements that outperform the baseline approaches, providing an effective solution for addressing unseen languages in the multilingual ASR task.

Instruction-based speech processing is becoming popular. Studies show that training with multiple tasks boosts performance, but collecting diverse, large-scale tasks and datasets is expensive. Thus, it is highly desirable to design a fundamental task that benefits other downstream tasks. This paper introduces a multi-talker speaking style captioning task to enhance the understanding of speaker and prosodic information. We used large language models to generate descriptions for multi-talker speech. Then, we trained our model with pre-training on this captioning task followed by instruction tuning. Evaluation on Dynamic-SUPERB shows our model outperforming the baseline pre-trained only on single-talker tasks, particularly in speaker and emotion recognition. The code and dataset are available at https://github.com/cyhuang-tw/speechcaps.

Recent advancements in large audio-language models (LALMs) have shown impressive capabilities in understanding and reasoning about audio and speech information. However, these models still face challenges, including hallucinating non-existent sound events, misidentifying the order of sound events, and incorrectly attributing sound sources, which undermine their reliability and real-world application. To systematically evaluate these issues, we propose three distinct tasks: object existence, temporal order, and object attribute within audio. These tasks assess the models’ comprehension of critical audio information aspects. Our experimental results reveal limitations in these fundamental tasks, underscoring the need for better models in recognizing specific sound events, determining event sequences, and identifying sound sources. To improve performance in these areas, we introduce a multi-turn chain-of-thought approach, which demonstrates significantly improved model performance across the proposed tasks.

Previous research has shown that the principal singular vectors of a pre-trained model’s weight matrices capture critical knowledge. In contrast, those associated with small singular values may contain noise or less reliable information. As a result, the LoRA-based parameter-efficient fine-tuning (PEFT) approach, which does not constrain the use of the spectral space, may not be effective for tasks that demand high representation capacity. In this study, we enhance existing PEFT techniques by incorporating the spectral information of pre-trained weight matrices into the fine-tuning process. We investigate spectral adaptation strategies with a particular focus on the additive adjustment of top singular vectors. This is accomplished by applying singular value decomposition (SVD) to the pre-trained weight matrices and restricting the fine-tuning within the top spectral space. Extensive speaker verification experiments on VoxCeleb1 and CN-Celeb1 demonstrate enhanced tuning performance with the proposed approach. Code is released at https://github.com/lizhepolyu/SpectralFT.

Speech Emotion Recognition (SER) is a crucial component in developing general-purpose AI agents capable of natural human-computer interaction. However, building robust multilingual SER systems remains challenging due to the scarcity of labeled data in languages other than English and Chinese. In this paper, we propose an approach to enhance SER performance in low SER resource languages by leveraging data from high-resource languages. Specifically, we employ expressive Speech-to-Speech translation (S2ST) combined with a novel bootstrapping data selection pipeline to generate labeled data in the target language. Extensive experiments demonstrate that our method is both effective and generalizable across different upstream models and languages. Our results suggest that this approach can facilitate the development of more scalable and robust multilingual SER systems. Our code is available at: https://github.com/hsi-che-lin/Improve-SER-via-S2ST

Recent end-to-end speech language models (SLMs) have expanded upon the capabilities of large language models (LLMs) by incorporating pre-trained speech models. However, these SLMs often undergo extensive speech instruction-tuning to bridge the gap between speech and text modalities. This requires significant annotation efforts and risks catastrophic forgetting of the original language capabilities. In this work, we present a simple yet effective automatic process for creating speech-text pair data that carefully injects speech paralinguistic understanding abilities into SLMs while preserving the inherent language capabilities of the text-based LLM. Our model demonstrates general capabilities for speech-related tasks without the need for speech instruction-tuning data, achieving impressive performance on Dynamic-SUPERB and AIR-Bench-Chat benchmarks. Furthermore, our model exhibits the ability to follow complex instructions derived from LLMs, such as specific output formatting and chain-of-thought reasoning. Our approach not only enhances the versatility and effectiveness of SLMs but also reduces reliance on extensive annotated datasets, paving the way for more efficient and capable speech understanding systems.1

Text-to-audio generation synthesizes realistic sounds or music given a natural language prompt. Diffusion-based frameworks, including the Tango and the AudioLDM series, represent the state-of-the-art in text-to-audio generation. Despite achieving high audio fidelity, they incur significant inference latency due to the slow diffusion sampling process. MAGNET, a mask-based model operating on discrete tokens, addresses slow inference through iterative mask-based parallel decoding. However, its audio quality still lags behind that of diffusion-based models. In this work, we introduce IMPACT, a text-to-audio generation framework that achieves high performance in audio quality and fidelity while ensuring fast inference. IMPACT utilizes iterative mask-based parallel decoding in a continuous latent space powered by diffusion modeling. This approach eliminates the fidelity constraints of discrete tokens while maintaining competitive inference speed. Results on AudioCaps demonstrate that IMPACT achieves state-of-the-art performance on key metrics including Fréchet Distance (FD) and Fréchet Audio Distance (FAD) while significantly reducing latency compared to prior models. The project website is available at https://audio-impact.github.io/.

Autoregressive next-token prediction with the Transformer decoder has become a de facto standard in large language models (LLMs), achieving remarkable success in Natural Language Processing (NLP) at scale. Extending this paradigm to audio poses unique challenges due to its inherently continuous nature. We research audio generation with a causal language model (LM) without discrete tokens. We leverage token-wise diffusion to model the continuous distribution of the next continuous-valued token. Our approach delivers significant improvements over previous discrete solution, AudioGen, achieving 20% and 40% relative gains on AudioCaps in Frechet Audio Distance (FAD) and Kullback-Leibler (KL) divergence, respectively. Additionally, we propose a novel masked next-token prediction task that incorporates masked prediction into the causal LM framework. On AudioCaps, the innovation yields 41% and 33% relative FAD improvements over AudioGen Base (285M) and AudioGen Large (1B) models, respectively, and is on par with the state-of-the-art (SOTA) diffusion models. Furthermore, we achieve these results with significantly fewer parameters—193M for our Base and 462M for our Large models.

Maintaining consistent model performance across domains is a fundamental challenge in machine learning. While recent work has explored using LLM-generated data for fine-tuning, its impact on cross-domain generalization remains poorly understood. This paper presents a systematic analysis revealing that fine-tuning with LLM-generated data not only improves target task performance but also reduces non-target task degradation compared to fine-tuning with ground truth data. Through analyzing the data sequence in tasks of various domains, we demonstrate that this enhancement of non-target task robustness stems from the reduction of high perplexity tokens found in LLM-generated sequences. Following our findings, we showed that masking high perplexity tokens in ground truth training data achieves similar non-target task performance preservation, comparable to using LLM-generated data. Extensive experiments across different model families and scales, including Gemma 2 IT 2B, Llama 3 8B Instruct, and three additional models, agree with our findings. To the best of our knowledge, this is the first work to provide an empirical explanation based on token perplexity reduction to mitigate catastrophic forgetting in LLMs after fine-tuning, offering valuable insights for developing more robust fine-tuning strategies.

Shih-Cheng Huang, Pin-Zu Li, Yu-chi Hsu, Kuang-Ming Chen, Yu Tung Lin, Shih-Kai Hsiao, Richard Tsai, Hung-yi Lee. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

In spoken dialogue, even if two current turns are the same sentence, their responses might still differ when they are spoken in different styles.The spoken styles, containing paralinguistic and prosodic information, mark the most significant difference between text and speech modality.When using text-only LLMs to model spoken dialogue, text-only LLMs cannot give different responses based on the speaking style of the current turn.In this paper, we focus on enabling LLMs to listen to the speaking styles and respond properly.Our goal is to teach the LLM that "even if the sentences are identical if they are spoken in different styles, their corresponding responses might be different".Since there is no suitable dataset for achieving this goal, we collect a speechto-speech dataset, StyleTalk, with the following desired characteristics: when two current speeches have the same content but are spoken in different styles, their responses will be different.To teach LLMs to understand and respond properly to the speaking styles, we propose the Spoken-LLM framework that can model the linguistic content and the speaking styles.We train Spoken-LLM using the StyleTalk dataset and devise a two-stage training pipeline to help the Spoken-LLM better learn the speaking styles.Based on extensive experiments, we show that Spoken-LLM outperforms text-only baselines and prior speech LLMs methods.

Using large language models (LLMs) for automatic evaluation has become an important evaluation method in NLP research.However, it is unclear whether these LLM-based evaluators can be applied in real-world classrooms to assess student assignments.This empirical report shares how we use GPT-4 as an automatic assignment evaluator in a university course with 1,028 students.Based on student responses, we find that LLM-based assignment evaluators are generally acceptable to students when students have free access to these LLM-based evaluators.However, students also noted that the LLM sometimes fails to adhere to the evaluation instructions.Additionally, we observe that students can easily manipulate the LLM-based evaluator to output specific strings, allowing them to achieve high scores without meeting the assignment rubric.Based on student feedback and our experience, we provide several recommendations for integrating LLM-based evaluators into future classrooms.Our observation also highlights potential directions for improving LLM-based evaluators, including their instruction-following ability and vulnerability to prompt hacking.

Deep Learning-based end-to-end Automatic Speech Recognition (ASR) has made significant strides but still struggles with performance on out-of-domain samples due to domain shifts in real-world scenarios.Test-Time Adaptation (TTA) methods address this issue by adapting models using test samples at inference time.However, current ASR TTA methods have largely focused on non-continual TTA, which limits cross-sample knowledge learning compared to continual TTA.In this work, we first propose a Fast-slow TTA framework for ASR that leverages the advantage of continual and non-continual TTA.Following this framework, we introduce Dynamic SUTA (DSUTA), an entropy-minimization-based continual TTA method for ASR.To enhance DSUTA robustness for time-varying multi-domain data, we design a dynamic reset strategy to automatically detect domain shifts and reset the model.Our method demonstrates superior performance on various noisy ASR datasets, outperforming both non-continual and continual TTA baselines while maintaining robustness to domain changes without requiring domain boundary information 1 .

Reinforcement learning from human feedback (RLHF) is a popular strategy for aligning large language models (LLMs) with desired behaviors.Reward modeling is a crucial step in RLHF.However, collecting paired preference data for training reward models is often costly and time-consuming, especially for domainspecific preferences requiring expert annotation.To address this challenge, we propose the Domain knowledge merged Reward Model (DogeRM), a novel framework that integrates domain-specific knowledge into a general reward model by model merging.The experiments demonstrate that DogeRM enhances performance across different benchmarks and provide a detailed analysis showcasing the effects of model merging, showing the great potential of facilitating model alignment.1

Synthetic data is widely used in speech recognition due to the availability of text-to-speech models, which facilitate adapting models to previously unseen text domains.However, existing methods suffer in performance when they finetune an automatic speech recognition (ASR) model on synthetic data as they suffer from the distributional shift commonly referred to as the synthetic-to-real gap.In this paper, we find that task arithmetic is effective at mitigating this gap.Our proposed method, SYN2REAL task vector, shows an average improvement of 10.03% improvement in word error rate over baselines on the SLURP dataset.Additionally, we show that an average of SYN2REAL task vectors, when we have real speeches from multiple different domains, can further adapt the original ASR model to perform better on the target text domain.

This study explores the proactive ability of LLMs to seek user support.We propose metrics to evaluate the trade-off between performance improvements and user burden, and investigate whether LLMs can determine when to request help under varying information availability.Our experiments show that without external feedback, many LLMs struggle to recognize their need for user support.The findings highlight the importance of external signals and provide insights for future research on improving support-seeking strategies.

Audio-visual synchronization aims to determine whether the mouth movements and speech in the video are synchronized. VocaLiST reaches state-of-the-art performance by incorporating multimodal Transformers to model audio-visual interact information. However, it requires high computing resources, making it impractical for real-world applications. This paper proposed an MTD-VocaLiST model, which is trained by our proposed multimodal Transformer distillation (MTD) loss. MTD loss enables MTDVocaLiST model to deeply mimic the cross-attention distribution and value-relation in the Transformer of VocaLiST. Additionally, we harness uncertainty weighting to fully exploit the interaction information across all layers. Our proposed method is effective in two aspects: From the distillation method perspective, MTD loss outperforms other strong distillation baselines. From the distilled model’s performance perspective: 1) MTDVocaLiST outperforms similar-size SOTA models, SyncNet, and Perfect Match models by 15.65% and 3.35%; 2) MTDVocaLiST reduces the model size of VocaLiST by 83.52%, yet still maintaining similar performance.

In the realm of spoken language understanding (SLU). numerous natural language understanding (NLU) methodologies have been adapted by supplying large language models (LLMs) with transcribed speech instead of conventional written text. In real-world scenarios, prior to input into an LLM. an automated speech recognition (ASR) system generates an output transcript hypothesis, where inherent errors can degrade subsequent SLU tasks. Here we introduce a method that utilizes the ASR system's lattice output instead of relying solely on the top hypothesis, aiming to encapsulate speech ambiguities and enhance SLU outcomes. Our in-context learning experiments, covering spoken question answering and intent classification. underline the LLM's resilience to noisy speech transcripts with the help of word confusion networks from lattices, bridging the SLU performance gap between using the top ASR hypothesis and an oracle upper bound. Additionally, we delve into the LLM's robustness to varying ASR performance conditions and scrutinize the aspects of in-context learning which prove the most influential.

Text language models have shown remarkable zero-shot capability in generalizing to unseen tasks when provided with well-formulated instructions. However, existing studies in speech processing primarily focus on limited or specific tasks. Moreover, the lack of standardized benchmarks hinders a fair comparison across different approaches. Thus, we present Dynamic-SUPERB, a benchmark designed for building universal speech models capable of leveraging instruction tuning to perform multiple tasks in a zero-shot fashion. To achieve comprehensive coverage of diverse speech tasks and harness instruction tuning, we invite the community to collaborate and contribute, facilitating the dynamic growth of the benchmark. To initiate, Dynamic-SUPERB features 55 evaluation instances by combining 33 tasks and 22 datasets. This spans a broad spectrum of dimensions, providing a comprehensive platform for evaluation. Additionally, we propose several approaches to establish benchmark baselines. These include the utilization of speech models, text language models, and the multimodal encoder. Evaluation results indicate that while these baselines perform reasonably on seen tasks, they struggle with unseen ones. We release all materials to the public and welcome researchers to collaborate on the project, advancing technologies in the field together1.

We introduce a new zero resource code-switched speech bench-mark designed to assess the code-switching capabilities of self-supervised speech encoders directly. We showcase a baseline system of language modeling on discrete units to demonstrate how the code-switching abilities of speech encoders can be assessed in a zero-resource manner. Our experiments encompass a variety of well-known speech encoders, including Wav2vec 2.0, HuBERT, XLSR, etc., on three tracks of different code-switched language pairs: Spanish-English, French-English, and Chinese-English. We examine the impact of pre-training languages and model size on benchmark performance. Notably, though our results demonstrate that speech encoders with multilingual pre-training, exemplified by XLSR, outperform monolingual variants (Wav2vec 2.0, HuBERT) in code-switching scenarios, there is still substantial room for improvement in their code-switching linguistic abilities.

Spoken Question Answering (SQA) is essential for machines to reply to user’s question by finding the answer span within a given spoken passage. SQA has been previously achieved without ASR to avoid recognition errors and Out-of-Vocabulary (OOV) problems. However, the real-world problem of Open-domain SQA (openSQA), in which the machine needs to first retrieve passages that possibly contain the answer from a spoken archive in addition, was never considered. This paper proposes the first known end-to-end frame-work, Speech Dense Passage Retriever (SpeechDPR), for the retrieval component of the openSQA problem. SpeechDPR learns a sentence-level semantic representation by distilling knowledge from the cascading model of unsupervised ASR (UASR) and text dense retriever (TDR). No manually transcribed speech data is needed. Initial experiments showed performance comparable to the cascading model of UASR and TDR, and significantly better when UASR was poor, verifying this approach is more robust to speech recognition errors.

Large Language Models (LLMs) have demonstrated superior abilities in tasks such as chatting, reasoning, and question-answering. However, standard LLMs may ignore crucial paralinguistic information, such as sentiment, emotion, and speaking style, which are essential for achieving natural, human-like spoken conversation, especially when such information is conveyed by acoustic cues. We therefore propose Paralinguistics-enhanced Generative Pretrained Transformer (ParalinGPT), an LLM that utilizes text and speech modalities to better model the linguistic content and paralinguistic attributes of spoken dialogue. The model takes the conversational context of text, speech embeddings, and paralinguistic attributes as input prompts within a serialized multitasking multimodal framework. Specifically, our framework serializes tasks in the order of current paralinguistic attribute prediction, response paralinguistic attribute prediction, and response text generation with autoregressive conditioning. We utilize the Switchboard-1 corpus, including its sentiment labels as the paralinguistic attribute, as our spoken dialogue dataset. Experimental results indicate the proposed serialized multitasking method outperforms typical sequence classification techniques on current and response sentiment classification. Furthermore, leveraging conversational context and speech embeddings significantly improves both response text generation and sentiment prediction. Our proposed framework achieves relative improvements of 6.7%, 12.0%, and 3.5% in current sentiment accuracy, response sentiment accuracy, and response text BLEU score, respectively.

Audio-visual representation learning aims to develop systems with human-like perception by utilizing correlation between auditory and visual information. However, current models often focus on a limited set of tasks, and generalization abilities of learned representations are unclear. To this end, we propose the AV-SUPERB benchmark that enables general-purpose evaluation of unimodal audio/visual and bimodal fusion representations on 7 datasets covering 5 audio-visual tasks in speech and audio processing. We evaluate 5 recent self-supervised models and show that none of these models generalize to all tasks, emphasizing the need for future study on improving universal model performance. In addition, we show that representations may be improved with intermediate-task fine-tuning and audio event classification with AudioSet serves as a strong intermediate task. We release our benchmark with evaluation code1and a model submission platform2to encourage further research in audio-visual learning.

Automatic speaker verification (ASV) is highly susceptible to adversarial attacks. Purification modules are usually adopted as a pre-processing to mitigate adversarial noise. However, they are commonly implemented across diverse experimental settings, rendering direct comparisons challenging. This paper comprehensively compares mainstream purification techniques in a unified framework. We find these methods often face a trade-off between user experience and security, as they struggle to simultaneously maintain genuine sample performance and reduce adversarial perturbations. To address this challenge, some efforts have extended purification modules to encompass detection capabilities, aiming to alleviate the trade-off. However, advanced purification modules will always come into the stage to surpass previous detection method. As a result, we further propose an easy-to-follow ensemble approach that integrates advanced purification modules for detection, achieving state-of-the-art (SOTA) performance in countering adversarial noise. Our ensemble method has great potential due to its compatibility with future advanced purification techniques.

Interspeech 2024, 1-5 September 2024, Kos, Greece

The diffusion model, a new generative modeling paradigm, has achieved significant success in generating images, audio, video, and text. It has been adapted for sequence-to-sequence text generation (Seq2Seq) through DiffuSeq, termed the S2S-Diffusion model. Existing S2S-Diffusion models predominantly rely on fixed or hand-crafted rules to schedule noise during the diffusion and denoising processes. However, these models are limited by non-contextualized noise, which fails to fully consider the characteristics of Seq2Seq tasks. In this paper, we propose the Meta-Diffu$B$ framework—a novel scheduler-exploiter S2S-Diffusion paradigm designed to overcome the limitations of existing S2S-Diffusion models. We employ Meta-Exploration to train an additional scheduler model dedicated to scheduling contextualized noise for each sentence. Our exploiter model, an S2S-Diffusion model, leverages the noise scheduled by our scheduler model for updating and generation. Meta-Diffu$B$ achieves state-of-the-art performance compared to previous S2S-Diffusion models and fine-tuned pre-trained language models (PLMs) across four Seq2Seq benchmark datasets. We further investigate and visualize the impact of Meta-Diffu$B$'s noise scheduling on the generation of sentences with varying difficulties. Additionally, our scheduler model can function as a "plug-and-play" model to enhance DiffuSeq without the need for fine-tuning during the inference stage.

Unsupervised automatic speech recognition (ASR) aims to learn the mapping between the speech signal and its corresponding textual transcription without the supervision of paired speech-text data. A word/phoneme in the speech signal is represented by a segment of speech signal with variable length and unknown boundary, and this segmental structure makes learning the mapping between speech and text challenging, especially without paired data. In this paper, we propose REBORN, Reinforcement-Learned Boundary Segmentation with Iterative Training for Unsupervised ASR. REBORN alternates between (1) training a segmentation model that predicts the boundaries of the segmental structures in speech signals and (2) training the phoneme prediction model, whose input is a segmental structure segmented by the segmentation model, to predict a phoneme transcription. Since supervised data for training the segmentation model is not available, we use reinforcement learning to train the segmentation model to favor segmentations that yield phoneme sequence predictions with a lower perplexity. We conduct extensive experiments and find that under the same setting, REBORN outperforms all prior unsupervised ASR models on LibriSpeech, TIMIT, and five non-English languages in Multilingual LibriSpeech. We comprehensively analyze why the boundaries learned by REBORN improve the unsupervised ASR performance.

Recent works have shown that large language model (LLM) agents are able to improve themselves from experience, which is an important ability for continuous enhancement post-deployment. However, existing benchmarks primarily evaluate their innate capabilities and do not assess their ability to improve over time. To address this gap, we introduce StreamBench, a pioneering benchmark designed to evaluate the continuous improvement of LLM agents over an input-feedback sequence. StreamBench simulates an online learning environment where LLMs receive a continuous flow of feedback stream and iteratively enhance their performance. In addition, we propose several simple yet effective baselines for improving LLMs on StreamBench, and provide a comprehensive analysis to identify critical components that contribute to successful streaming strategies. Our work serves as a stepping stone towards developing effective online learning strategies for LLMs, paving the way for more adaptive AI systems in streaming scenarios.

Speech Emotion Recognition (SER) faces a distinct challenge compared to other speech-related tasks because the annotations will show the subjective emotional perceptions of different annotators. Previous SER studies often view the subjectivity of emotion perception as noise by using the majority rule or plurality rule to obtain the consensus labels. However, these standard approaches overlook the valuable information of labels that do not agree with the consensus and make it easier for the test set. Emotion perception can have co-occurring emotions in realistic conditions, and it is unnecessary to regard the disagreement between raters as noise. To bridge the SER into a multi-label task, we introduced an “all-inclusive rule,” which considers all available data, ratings, and distributional labels as multi-label targets and a complete test set. We demonstrated that models trained with multi-label targets generated by the proposed AR outperform conventional single-label methods across incomplete and complete test sets.

Mainstream zero-shot TTS production systems like Voicebox and Seed-TTS achieve human parity speech by leveraging Flow-matching and Diffusion models, respectively. Unfortunately, human-level audio synthesis leads to identity misuse and information security issues. Currently, many anti-spoofing models have been developed against deepfake audio. However, the efficacy of current state-of-the-art anti-spoofing models in countering audio synthesized by diffusion and flow-matching based TTS systems remains unknown. In this paper, we proposed the Diffusion and Flow-matching based Audio Deepfake (DFADD) dataset. The DFADD dataset collected the deepfake audio based on advanced diffusion and flowmatching TTS models. Additionally, we reveal that current anti-spoofing models lack sufficient robustness against highly human-like audio generated by diffusion and flow-matching TTS systems. The proposed DFADD dataset addresses this gap and provides a valuable resource for developing more resilient anti-spoofing models.

Neural speech editing advancements have raised concerns about their misuse in spoofing attacks. Traditional partially edited speech corpora primarily focus on cut-and-paste edits, which, while maintaining speaker consistency, often introduce detectable discontinuities. Recent methods, like $\mathrm{A}^{3} \mathrm{~T}$ and Voicebox, improve transitions by leveraging contextual information. To foster spoofing detection research, we introduce the Speech INfilling Edit (SINE) dataset, created with Voicebox. We detailed the process of re-implementing Voicebox training and dataset creation. Subjective evaluations confirm that speech edited using this novel technique is more challenging to detect than conventional cut-and-paste methods. Despite human difficulty, experimental results demonstrate that self-supervised-based detectors can achieve remarkable performance in detection, localization, and generalization across different edit methods. The dataset and related models will be made available at: https://jasonswfu.github.io/SINE_dataset/index.html

In this work, we introduce Speech-Copilot, a modular framework for instruction-oriented speech-processing tasks that minimizes human effort in toolset construction. Unlike end-to-end methods using large audio-language models, Speech-Copilot builds speech processing-specific toolsets by analyzing pre-collected task instructions and breaking tasks into manageable sub-tasks. It features a flexible agent that performs tasks through program generation based on large language models. Our approach achieves state-of the-art performance on the Dynamic-SUPERB benchmark, demonstrating its effectiveness across diverse speech-processing tasks. Key contributions include: 1) developing an innovative framework for speech processing-specific toolset construction, 2) establishing a high-performing agent based on large language models, and 3) offering a new perspective on addressing challenging instruction-oriented speech-processing tasks. Without additional training required by end-to-end approaches, our method provides a flexible and extendable solution for a wide range of speech-processing applications.

Warning: This paper may contain texts with uncomfortable content.Large Language Models (LLMs) have achieved remarkable performance in various tasks, including those involving multimodal data like speech. However, these models often exhibit biases due to the nature of their training data. Recently, more Speech Large Language Models (SLLMs) have emerged, underscoring the urgent need to address these biases. This study introduces Spoken Stereoset, a dataset specifically designed to evaluate social biases in SLLMs. By examining how different models respond to speech from diverse demographic groups, we aim to identify these biases. Our experiments reveal significant insights into their performance and bias levels. The findings indicate that while most models show minimal bias, some still exhibit slightly stereotypical or anti-stereotypical tendencies.

There have been many studies on analyzing self-supervised speech Transformers, in particular, with layer-wise analysis. It is, however, desirable to have an approach that can pinpoint exactly a subset of neurons that is responsible for a particular property of speech, being amenable to model pruning and model editing. In this work, we identify a set of property neurons in the feedforward layers of Transformers to study how speech-related properties, such as phones, gender, and pitch, are stored. When removing neurons of a particular property (a simple form of model editing), the respective downstream performance significantly degrades, showing the importance of the property neurons. We apply this approach to pruning the feedforward layers in Transformers, where most of the model parameters are. We show that protecting property neurons during pruning is significantly more effective than normbased pruning.

Speech Integrated Large Language Models (SILLMs) combine large language models with speech perception to perform diverse tasks, such as emotion recognition to speaker verification, demonstrating universal audio understanding capability. However, these models may amplify biases present in training data, potentially leading to biased access to information for marginalized groups. This work introduces a curated spoken bias evaluation toolkit and corresponding dataset. We evaluate gender bias in SILLMs across four semantic-related tasks: speech-to-text translation (STT), spoken coreference resolution (SCR), spoken sentence continuation (SSC), and spoken question answering (SQA). Our analysis reveals that bias levels are language-dependent and vary with different evaluation methods. Our findings emphasize the necessity of employing multiple approaches to comprehensively assess biases in SILLMs, providing insights for developing fairer SILLM systems.

Despite their impressive success, training foundation models remains computationally costly. This paper investigates how to efficiently train speech foundation models with self-supervised learning (SSL) under a limited compute budget. We examine critical factors in SSL that impact the budget, including model architecture, model size, and data size. Our goal is to make analytical steps toward understanding the training dynamics of speech foundation models. We benchmark SSL objectives in an entirely comparable setting and find that other factors contribute more significantly to the success of SSL. Our results show that slimmer model architectures outperform common small architectures under the same compute and parameter budget. We demonstrate that the size of the pre-training data remains crucial, even with data augmentation during SSL training, as performance suffers when iterating over limited data. Finally, we identify a trade-off between model size and data size, highlighting an optimal model size for a given compute budget.

Recent advances in automatic speech recognition (ASR) often rely on large speech foundation models for generating high-quality transcriptions. However, these models can be impractical due to limited computing resources. The situation is even more severe in terms of more realistic or difficult scenarios, such as code-switching ASR (CS-ASR). To address this, we present a framework for developing more efficient models for CS-ASR through knowledge distillation using realistic speech-only data. Our proposed method, Leave No Knowledge Behind During Knowledge Distillation ($\mathrm{K}^{2} \mathrm{D}$), leverages both the teacher model’s knowledge and additional insights from a small auxiliary model. We evaluate our approach on two in-domain and two out-domain datasets, demonstrating that $\mathrm{K}^{2} \mathrm{D}$ is effective. By conducting $\mathrm{K}^{2} \mathrm{D}$ on the unlabeled realistic data, we have successfully obtained a 2 -time smaller model with 5 -time faster generation speed while outperforming the baseline methods and the teacher model on all the testing sets.

Self-supervised learning (SSL) models have shown exceptional capabilities across various speech-processing tasks. Continuous SSL representations are effective but suffer from high computational and storage demands. On the other hand, discrete SSL representations, although with degraded performance, reduce transmission and storage costs, and improve input sequence efficiency through de-duplication and subword-modeling. To boost the performance of discrete representations for ASR, we introduce a novel fusion mechanism that integrates two discrete representations. The fusion mechanism preserves all the benefits of discrete representation while enhancing the model’s performance by integrating complementary information. Additionally, we explore “self-augmented” discrete representations, which apply transformations to a single continuous SSL representation, eliminating the fusion mechanism’s dependency on multiple SSL models and further decreasing its inference costs. Experimental results on benchmarks, including LibriSpeech and ML-SUPERB, indicate up to 19% and 24% relative character error rate improvement compared with the non-fusion baseline, validating the effectiveness of our proposed methods.

Speech emotion recognition (SER) is an essential technology for human-computer interaction systems. However, the previous study reveals that 80.77% of SER papers yield results that cannot be reproduced on the well-known IEMOCAP dataset. The main reason for reproducibility challenges is that the database did not provide standard data splits (e.g., train, development, and test sets). Prior papers could define its partition, but they did not provide details of the partition or source code for processing the partition. Therefore, this work aims to make SER open and reproducible to everyone. We develop the EMO-SUPERB, shorted for EMOtion Speech Universal PERformance Benchmark, including a user-friendly codebase to leverage 16 state-of-the-art (SOTA) speech self-supervised learning models for exhaustive evaluation plus one SOTA SER model across 6 open-source SER datasets in English and Chinese. We make all resources open-source to facilitate future developments in SER. Researchers can easily upload their systems or datasets to EMO-SUPERB, and we name the project “Open-Emotion”.

Neural audio codec models are becoming increasingly important as they serve as tokenizers for audio, enabling efficient transmission or facilitating speech language modeling. The ideal neural audio codec should maintain content, paralinguistics, speaker characteristics, and audio information even at low bitrates. Recently, numerous advanced neural codec models have been proposed. However, codec models are often tested under varying experimental conditions. As a result, we introduce the Codec-SUPERB challenge at SLT 20241, designed to facilitate fair and lightweight comparisons among existing codec models and inspire advancements in the field. This challenge brings together representative speech applications and objective metrics, and carefully selects license-free datasets, sampling them into small sets to reduce evaluation computation costs. This paper presents the challenge’s rules, datasets, participant systems, results, and findings.1https://codecsuperb.github.io/

This research explores how the information of prompts interacts with the high-performing speech recognition model, Whisper. We compare its performances when prompted by prompts with correct information and those corrupted with incorrect information. Our results unexpectedly show that Whisper may not understand the textual prompts in a human-expected way. Additionally, we find that performance improvement is not guaranteed even with stronger adherence to the topic information in textual prompts. It is also noted that English prompts generally outperform Mandarin ones on datasets of both languages, likely due to differences in training data distributions for these languages despite the mismatch with pre-training scenarios. Conversely, we discover that Whisper exhibits awareness of misleading information in language tokens by ignoring incorrect language tokens and focusing on the correct ones. In sum, We raise insightful questions about Whisper’s prompt understanding and reveal its counter-intuitive behaviors. We encourage further studies.

Prompting has become a practical method for utilizing pre-trained language models (LMs). This approach offers several advantages. It allows an LM to adapt to new tasks with minimal training and parameter updates, thus achieving efficiency in both storage and computation. Additionally, prompting modifies only the LM's inputs and harnesses the generative capabilities of language models to address various downstream tasks in a unified manner. This significantly reduces the need for human labor in designing task-specific models. These advantages become even more evident as the number of tasks served by the LM scales up. Motivated by the strengths of prompting, we are the first to explore the potential of prompting speech LMs in the domain of speech processing. Recently, there has been a growing interest in converting speech into discrete units for language modeling. Our pioneer research demonstrates that these quantized speech units are highly versatile within our unified prompting framework. Not only can they serve as class labels, but they also contain rich phonetic information that can be re-synthesized back into speech signals for speech generation tasks. Specifically, we reformulate speech processing tasks into speech-to-unit generation tasks. As a result, we can seamlessly integrate tasks such as speech classification, sequence generation, and speech generation within a single, unified prompting framework. The experiment results show that the prompting method can achieve competitive performance compared to the strong fine-tuning method based on self-supervised learning models with a similar number of trainable parameters. The prompting method also shows promising results in the few-shot setting. Moreover, with the advanced speech LMs coming into the stage, the proposed prompting framework attains great potential.

Non-autoregressive approaches, especially those that generate output in a one-pass forward manner, have shown great potential in improving the inference speed of translation models. However, these approaches often suffer from a significant drop in translation quality compared to autoregressive models (AT). To tackle this challenge, this paper introduces a series of innovative techniques to enhance the translation quality of non-autoregressive neural machine translation (NAT) models while still maintaining a substantial acceleration in inference speed. Specifically, we propose a method called CTCPMLM, which involves fine-tuning Pretrained Multilingual Language Models (PMLMs) with the Connectionist Temporal Classification (CTC) loss to effectively train NAT models. Additionally, we adopt the MASK insertion scheme instead of token duplication for up-sampling and present an embedding distillation method to further enhance the performance of NAT models. In our experiments, CTCPMLM surpasses the performance of the baseline autoregressive model (Transformerbase) on various datasets, including WMT'14 DE$\leftrightarrow$EN, WMT'16 RO$\leftrightarrow$EN, and IWSLT'14 DE$\leftrightarrow$EN. Moreover, CTCPMLM represents the current state-of-the-art among NAT models. Notably, our model achieves superior results compared to the baseline autoregressive model on the IWSLT'14 En$\leftrightarrow$De and WMT'16 En$\leftrightarrow$Ro datasets, even without using distillation data during training. Particularly, on the IWSLT'14 DE$\rightarrow$EN dataset, our model achieves an impressive BLEU score of 39.93, surpassing AT models and establishing a new state-of-the-art. Additionally, our model exhibits a remarkable speed improvement of 16.35 times compared to the autoregressive model.

The foundation model paradigm leverages a shared foundation model to achieve state-of-the-art (SOTA) performance for various tasks, requiring minimal downstream-specific data collection and modeling. This approach has proven crucial in the field of Natural Language Processing (NLP). However, the speech processing community lacks a similar setup to explore the paradigm systematically. To bridge this gap, we establish the Speech processing Universal PERformance Benchmark (SUPERB). SUPERB represents an ecosystem designed to evaluate foundation models across a wide range of speech processing tasks, facilitating the sharing of results on an online leaderboard and fostering collaboration through a community-driven benchmark database that aids in new development cycles. We present a unified learning framework for solving the speech processing tasks in SUPERB with the frozen foundation model followed by task-specialized lightweight prediction heads. Combining our results with community submissions, we verify that the framework is simple yet effective, as the best-performing foundation model shows competitive generalizability across most SUPERB tasks. Finally, we conduct a series of analyses to offer an in-depth understanding of SUPERB and speech foundation models, including information flows across tasks inside the models and the statistical significance and robustness of the benchmark.

Human evaluation is indispensable and inevitable for assessing the quality of texts generated by machine learning models or written by humans.However, human evaluation is very difficult to reproduce and its quality is notoriously unstable, hindering fair comparisons among different natural language processing (NLP) models and algorithms.Recently, large language models (LLMs) have demonstrated exceptional performance on unseen tasks when only the task instructions are provided.In this paper, we explore if such an ability of the LLMs can be used as an alternative to human evaluation.We present the LLMs with the exact same instructions, samples to be evaluated, and questions used to conduct human evaluation, and then ask the LLMs to generate responses to those questions; we dub this LLM evaluation.We use human evaluation and LLM evaluation to evaluate the texts in two NLP tasks: openended story generation and adversarial attacks.We show that the result of LLM evaluation is consistent with the results obtained by expert human evaluation: the texts rated higher by human experts are also rated higher by the LLMs.We also find that the results of LLM evaluation are stable over different formatting of the task instructions and the sampling algorithm used to generate the answer.We are the first to show the potential of using LLMs to assess the quality of texts and discuss the limitations and ethical considerations of LLM evaluation.

Suwon Shon, Siddhant Arora, Chyi-Jiunn Lin, Ankita Pasad, Felix Wu, Roshan S Sharma, Wei-Lun Wu, Hung-yi Lee, Karen Livescu, Shinji Watanabe. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Derek Xu, Shuyan Dong, Changhan Wang, Suyoun Kim, Zhaojiang Lin, Bing Liu, Akshat Shrivastava, Shang-Wen Li, Liang-Hsuan Tseng, Guan-Ting Lin, Alexei Baevski, Hung-yi Lee, Yizhou Sun, Wei Wang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Prompting and adapter tuning have emerged as efficient alternatives to fine-tuning (FT) methods. However, existing studies on speech prompting focused on classification tasks and failed on more complex sequence generation tasks. Besides, adapter tuning is primarily applied with a focus on encoder-only self-supervised models. Our experiments show that prompting on Wav2Seq, a self-supervised encoder-decoder model, surpasses previous works in sequence generation tasks. It achieves a remarkable 53% relative improvement in word error rate for ASR and a 27% in F1 score for slot filling. Additionally, prompting competes with the FT method in the low-resource scenario. Moreover, we show the transferability of prompting and adapter tuning on Wav2Seq in cross-lingual ASR. When limited trainable parameters are involved, prompting and adapter tuning consistently outperform conventional FT across 7 languages. Notably, in the low-resource scenario, prompting consistently outperforms adapter tuning.

This paper presents an effective transfer learning framework for language adaptation in text-to-speech systems, with a focus on achieving language adaptation using minimal labeled and unlabeled data. While many works focus on reducing the usage of labeled data, very few consider minimizing the usage of unlabeled data. By utilizing self-supervised features in the pretraining stage, replacing the noisy portion of pseudo labels with these features during fine-tuning, and incorporating an embedding initialization trick, our method leverages more information from unlabeled data compared to conventional approaches. Experimental results show that our framework is able to synthesize intelligible speech in unseen languages with only 4 utterances of labeled data and 15 minutes of unlabeled data. Our methodology continues to surpass conventional techniques, even when a greater volume of data is accessible. These findings highlight the potential of our data-efficient language adaptation framework.

This paper introduces a novel voice conversion (VC) model, guided by text instructions such as “articulate slowly with a deep tone“ or “speak in a cheerful boyish voice”. Unlike traditional methods that rely on reference utterances to determine the attributes of the converted speech, using text instruction adds versatility and specificity to voice conversion. The proposed VC model is a neural codec language model which processes a sequence of discrete codes, resulting in the code sequence of converted speech. It utilizes text instructions as style prompts to modify the prosody and emotional information of the given speech. In contrast to previous approaches, which often rely on employing separate encoders like prosody and content encoders to handle different aspects of the source speech, our model handles various information of speech in an end-to-end manner. Experiments have demonstrated the impressive capabilities of our model in comprehending instructions and delivering reasonable results1.

Self-supervised models have had great success in learning speech representations that can generalize to various downstream tasks. However, most self-supervised models require a large amount of compute and multiple GPUs to train, significantly hampering the development of self-supervised learning. In an attempt to reduce the computation of training, we revisit the training of HuBERT, a highly successful self-supervised model. We improve and simplify several key components, including the loss function, input representation, and training in multiple stages. Our model, MelHuBERT, is able to achieve favorable performance on phone recognition, speaker identification, and automatic speech recognition against HuBERT, while saving 31.2% of the pre-training time, or equivalently 33.5% MACs per one second speech. The code and pretrained models are available in https://github.com/nervjack2/MelHuBERT.

The 2023 Multilingual Speech Universal Performance Benchmark (ML-SUPERB) Challenge expands upon the acclaimed SUPERB framework, emphasizing self-supervised models in multilingual speech recognition and language identification. The challenge comprises a research track focused on applying ML-SUPERB to specific multilingual subjects, a Challenge Track for model submissions, and a New Language Track where language resource researchers can contribute and evaluate their low-resource language data in the context of the latest progress in multilingual speech recognition. The challenge garnered 12 model submissions and 54 language corpora, resulting in a comprehensive benchmark encompassing 154 languages. The findings indicate that merely scaling models is not the definitive solution for multilingual speech tasks, and a variety of speech/voice types present significant challenges in multilingual speech processing.

SUPERB was proposed to evaluate the generalizability of self-supervised learning (SSL) speech models across various tasks. However, it incurs high computational costs due to the large datasets and diverse tasks. In this paper, we introduce MiniSUPERB, a lightweight benchmark that efficiently evaluates SSL speech models with comparable results to SUPERB but lower computational costs significantly. We carefully select representative tasks, sample datasets, and extract model representations offline. Our approach achieves a Spearman’s rank correlation of 0.954 and 0.982 with SUPERB Paper and SUPERB Challenge, respectively. Additionally, we reduce the computational cost by 97 % in terms of Multiply-ACcumulate operations (MACs). Furthermore, we evaluate SSL speech models in few-shot scenarios and observe significant variations in their performance. To our knowledge, this is the first study to examine both the computational cost of the model itself and the cost of evaluating it on a benchmark.11Our code is available at https://github.com/Comet0322/MiniSUPERB

Zero-shot singing voice synthesis (SVS), the task to synthesize the singing voice of an arbitrary target singer, has gained increasing attentions in the past few years. Several recently proposed systems have demonstrated promising results on this task. However, these systems require detailed musical features at the frame level as the musical content. To deal with this issue, we propose a model that performs zero-shot SVS with only musical score as the musical content condition. To help model training, we build an acoustic encoder that extracts linguistic features from audio, and train it with the lyrics transcription objective. The output of the acoustic encoder serves as an alternative to the musical score, allowing the SVS model to learn from weakly labeled data. Results suggest that the proposed method outperforms baseline semi-supervised method in both subjective and objective tests.

This work investigates the use of large-scale, English-only pre-trained models (CLIP and HuBERT) for multilingual image-speech retrieval. For non-English image-speech retrieval, we outperform the current state-of-the-art performance by a wide margin both when training separate models for each language, and with a single model which processes speech in all three languages. We identify key differences in model behavior and performance between English and non-English settings, attributable to the English-only pre-training of CLIP and HuBERT, and investigate how fine-tuning the pre-trained models impacts these differences. Finally, we show that our models can be used for mono- and cross-lingual speech-text retrieval and cross-lingual speech-speech retrieval, despite never having seen any parallel speech-text or speech-speech data during training.

The sequence length along the time axis is often the dominant factor of the computation in speech processing. Works have been proposed to reduce the sequence length for lowering the computational cost in self-supervised speech models. However, different downstream tasks have different tolerance of sequence compressing, so a model that produces a fixed compressing rate may not fit all tasks. In this work, we introduce a once-for-all (OFA) sequence compression framework for self-supervised speech models that supports a continuous range of operating compressing rates. The framework is evaluated on various tasks, showing marginal degradation compared to the fixed compressing rate variants with a smooth performance-efficiency trade-off. We further explore adaptive compressing rate learning, demonstrating the ability to select task-specific preferred frame periods without needing a grid search.

This paper describes the ESPnet Unsupervised ASR Open-source Toolkit (EURO), an end-to-end open-source toolkit for unsupervised automatic speech recognition (UASR). EURO adopts the state-of-the-art UASR learning method introduced by the Wav2vec-U, originally implemented at FAIRSEQ, which leverages self-supervised speech representations and adversarial training. In addition to wav2vec2, EURO extends the functionality and promotes reproducibility for UASR tasks by integrating S3PRL and k2, resulting in flexible frontends from 27 self-supervised models and various graph-based decoding strategies. EURO is implemented in ESPnet and follows its unified pipeline to provide UASR recipes with a complete setup. This improves the pipeline’s efficiency and allows EURO to be easily applied to existing datasets in ESPnet. Extensive experiments on three mainstream self-supervised models demonstrate the toolkit’s effectiveness and achieve state-of-the-art UASR performance on TIMIT and LibriSpeech datasets. EURO will be publicly available at https://github.com/espnet/espnet, aiming to promote this exciting and emerging research area based on UASR through open-source activity.

In Spoken language understanding (SLU), a natural solution is concatenating pre-trained speech models (e.g. HuBERT) and pretrained language models (PLM, e.g. T5). Most previous works use pre-trained language models with subword-based tokenization. However, the granularity of input units affects the alignment of speech model outputs and language model inputs, and PLM with character-based tokenization is underexplored. In this work, we conduct extensive studies on how PLMs with different tokenization strategies affect spoken language understanding task including spoken question answering (SQA) and speech translation (ST).We further extend the idea to create T5lephone1, a variant of T5 that is pretrained using phonemicized text. We initialize T5lephone with existing PLMs to pretrain it using relatively lightweight computational resources. We reached state-of-the-art on NMSQA, and the T5lephone model exceeds T5 with other types of units on end-to-end SQA and ST. Our code is publicly available.2

Personalized TTS is an exciting and highly desired application that allows users to train their TTS voice using only a few recordings. However, TTS training typically requires many hours of recording and a large model, making it unsuitable for deployment on mobile devices. To overcome this limitation, related works typically require fine-tuning a pre-trained TTS model to preserve its ability to generate high-quality audio samples while adapting to the target speaker’s voice. This process is commonly referred to as "voice cloning." Although related works have achieved significant success in changing the TTS model’s voice, they are still required to fine-tune from a large pre-trained model, resulting in a significant size for the voice-cloned model. In this paper, we propose applying trainable structured pruning to voice cloning. By training the structured pruning masks with voice-cloning data, we can produce a unique pruned model for each target speaker. Our experiments demonstrate that using learnable structured pruning, we can compress the model size to 7 times smaller while achieving comparable voice-cloning performance.

Distilled self-supervised models have shown competitive performance and efficiency in recent years. However, there is a lack of experience in jointly distilling multiple self-supervised speech models. In our work, we performed Ensemble Knowledge Distillation (EKD) on various self-supervised speech models such as HuBERT, RobustHuBERT, and WavLM. We tried two different aggregation techniques, layerwise-average and layerwise-concatenation, to the representations of different teacher models and found that the former was more effective. On top of that, we proposed a multiple prediction head method for student models to predict different layer outputs of multiple teacher models simultaneously. The experimental results show that our method improves the performance of the distilled models on four downstream speech processing tasks, Phoneme Recognition, Speaker Identification, Emotion Recognition, and Automatic Speech Recognition in the hidden-set track of the SUPERB benchmark.

Spoken language understanding (SLU) is a task aiming to extract high-level semantics from spoken utterances. Previous works have investigated the use of speech self-supervised models and textual pre-trained models, which have shown reasonable improvements to various SLU tasks. However, because of the mismatched modalities between speech signals and text tokens, previous methods usually need complex designs of the frameworks. This work proposes a simple yet efficient unsupervised paradigm that connects speech and textual pre-trained models, resulting in an unsupervised speech-to-semantic pre-trained model for various tasks in SLU. To be specific, we propose to use unsupervised automatic speech recognition (ASR) as a connector that bridges different modalities used in speech and textual pre-trained models. Our experiments show that unsupervised ASR itself can improve the representations from speech self-supervised models. More importantly, it is shown as an efficient connector between speech and textual pre-trained models, improving the performances of five different SLU tasks. Notably, on spoken question answering, we reach the state-of-the-art result over the challenging NMSQA benchmark.

Past work on unsupervised parsing is constrained to written form. In this paper, we present the first study on unsupervised spoken constituency parsing given unlabeled spoken sentences and unpaired textual data. The goal is to determine the spoken sentences’ hierarchical syntactic structure in the form of constituency parse trees, such that each node is a span of audio that corresponds to a constituent. We compare two approaches: (1) cascading an unsupervised automatic speech recognition (ASR) model and an unsupervised parser to obtain parse trees on ASR transcripts, and (2) direct training an unsupervised parser on continuous word-level speech representations. This is done by first splitting utterances into sequences of word-level segments, and aggregating self-supervised speech representations within segments to obtain segment embeddings. We find that separately training a parser on the unpaired text and directly applying it on ASR transcripts for inference produces better results for unsupervised parsing. Additionally, our results suggest that accurate segmentation alone may be sufficient to parse spoken sentences accurately. Finally, we show the direct approach may learn head-directionality correctly for both head-initial and head-final languages without any explicit inductive bias.

Aiming to produce reinforcement learning (RL) policies that are human-interpretable and can generalize better to novel scenarios, Trivedi et al. (2021) present a method (LEAPS) that first learns a program embedding space to continuously parameterize diverse programs from a pre-generated program dataset, and then searches for a task-solving program in the learned program embedding space when given a task. Despite the encouraging results, the program policies that LEAPS can produce are limited by the distribution of the program dataset. Furthermore, during searching, LEAPS evaluates each candidate program solely based on its return, failing to precisely reward correct parts of programs and penalize incorrect parts. To address these issues, we propose to learn a meta-policy that composes a series of programs sampled from the learned program embedding space. By learning to compose programs, our proposed hierarchical programmatic reinforcement learning (HPRL) framework can produce program policies that describe out-of-distributionally complex behaviors and directly assign credits to programs that induce desired behaviors. The experimental results in the Karel domain show that our proposed framework outperforms baselines. The ablation studies confirm the limitations of LEAPS and justify our design choices.

In this work, we introduce a "score-based assessment" framework for estimating the transferability of pre-trained speech models (PSMs) for fine-tuning target tasks.We leverage upon two representation theories, Bayesian likelihood estimation and optimal transport, to generate rank scores for the PSM candidates using the extracted representations.Our framework efficiently computes transferability scores without actual finetuning of candidate models or layers by making a temporal independent hypothesis.We evaluate some popular supervised speech models (e.g., Conformer RNN-Transducer) and selfsupervised speech models (e.g., HuBERT) in cross-layer and cross-model settings using public data.Experimental results show a high Spearman's rank correlation and low p-value between our estimation framework and fine-tuning ground truth.Our proposed transferability framework requires less computational time and resources, making it a resource-saving and timeefficient approach for tuning speech foundation models.

Generative Adversarial Networks (GANs) have been popularly researched in natural language generation, so-called Language GANs. Existing works adopt reinforcement learning (RL) based methods such as policy gradients for training Language GANs. The previous research of Language GANs usually focuses on stabilizing policy gradients or applying robust architectures (such as the large-scale pre-trained GPT-2) to achieve better performance. However, the quality and diversity of sampling are not guaranteed simultaneously. In this article, we propose a novel meta-learning-based generative adversarial network, Meta Exploration GAN (MetaEx-GAN), for ensuring the quality and diversity of sampling (sampling efficiency). In the proposed MetaEx-GAN, we develop an explorer trained by Meta Exploration to sample from the generated data to achieve better sampling efficiency. MetaEx-GAN employs MetaEx first applied to Language GANs to achieve better performance. We also propose a critical training method for MetaEx-GAN on the NLG task. According to our experimental results, MetaEx-GAN achieves state-of-the-art performance compared with existing Language GANs methods. Our experiments also demonstrate the generality of MetaEx-GAN with different architectures (involving GPT-2) and how MetaEx-GAN operates to improve Language GANs.

Autoregressive neural vocoders have achieved outstanding performance and are widely used in speech synthesis tasks such as text-to-speech and voice conversion. An autoregressive vocoder predicts a sample at some time step conditioned on those at previous time steps. Though it can generate highly natural human speech, the iterative generation inevitably makes the synthesis time proportional to the utterance length, leading to low efficiency. Many works were dedicated to generating the whole speech time sequence in parallel and then proposed GAN-based, flow-based, and score-based vocoders. This paper proposed a new thought for the autoregressive generation. Instead of iteratively predicting samples in a time sequence, the proposed model performs frequency-wise autoregressive generation (FAR) and bit-wise autoregressive generation (BAR) to synthesize speech. In FAR, a speech utterance is first split into different frequency subbands. The proposed model generates a subband conditioned on the previously generated one. A full-band speech can then be reconstructed from these generated subbands. Similarly, in BAR, an 8-bit quantized signal is generated iteratively from the first bit. By redesigning the autoregressive method to compute in domains other than the time domain, the number of iterations in the proposed model is no longer proportional to the utterance's length but to the number of subbands/bits. The inference efficiency is hence significantly increased. Besides, a post-filter is employed to sample audio signals from output posteriors, and its training objective is designed based on the characteristics of the proposed autoregressive methods. The experimental results show that the proposed model can synthesize speech faster than real-time without GPU acceleration. Compared with the baseline autoregressive and non-autoregressive vocoders, the proposed model achieves better MUSHRA results and shows good generalization ability while synthesizing 44 kHz speech or utterances from unseen speakers.

Pre-training language models (LMs) on large-scale unlabeled text data makes the model much easier to achieve exceptional downstream performance than their counterparts directly trained on the downstream tasks. In this work, we study what specific traits in the pre-training data, other than the semantics, make a pre-trained LM superior to their counterparts trained from scratch on downstream tasks. We propose to use artificially constructed datasets as the pre-training data to exclude the effect of semantics, and further control what characteristics the pre-training corpora have. By fine-tuning the pre-trained models on GLUE benchmark, we can learn how beneficial it is to transfer the knowledge from the model trained on the dataset possessing that specific trait. We define and discuss three different characteristics in the artificial dataset: 1) matching the token's uni-gram or bi-gram distribution between pre-training and downstream fine-tuning, 2) the presence of the explicit dependencies among the tokens in a sequence, 3) the length of the implicit dependencies among the tokens in a sequence. Our experiments show that the explicit dependencies in the sequences of the pre-training data are critical to the downstream performance. Our results also reveal that models achieve better downstream performance when pre-trained on a dataset with a longer range of implicit dependencies. Based on our analysis, we find that models pre-trained with artificial datasets are prone to learn spurious correlation in downstream tasks. Our work reveals that even if the LMs are not pre-trained on natural language, they still gain transferability on certain human language downstream tasks once the LMs learn to model the token dependencies in the sequences. This result helps us understand the exceptional transferability of pre-trained LMs.

Hsiang-Sheng Tsai, Heng-Jui Chang, Wen-Chin Huang, Zili Huang, Kushal Lakhotia, Shu-wen Yang, Shuyan Dong, Andy Liu, Cheng-I Lai, Jiatong Shi, Xuankai Chang, Phil Hall, Hsuan-Jui Chen, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, Hung-yi Lee. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Self-supervised speech representation learning methods like wav2vec 2.0 and Hidden-unit BERT (HuBERT) leverage unlabeled speech data for pre-training and offer good representations for numerous speech processing tasks. Despite the success of these methods, they require large memory and high pre-training costs, making them inaccessible for researchers in academia and small companies. Therefore, this paper introduces DistilHuBERT, a novel multi-task learning framework to distill hidden representations from a HuBERT model directly. This method reduces HuBERT’s size by 75% and 73% faster while retaining most performance in ten different tasks. Moreover, DistilHuBERT required little training time and data, opening the possibilities of pre-training personal and on-device SSL models for speech.

Any-to-any voice conversion technologies convert the vocal timbre of an utterance to any speaker even unseen during training. Although there have been several state-of-the-art any-to-any voice conversion models, they were all based on clean utterances to convert successfully. However, in real-world scenarios, it is difficult to collect clean utterances of a speaker, and they are usually degraded by noises or reverberations. It thus becomes highly desired to understand how these degradations affect voice conversion and build a degradation-robust model. We report in this paper the first comprehensive study on the degradation robustness of any-to-any voice conversion. We show that the performance of state-of-the-art models nowadays was severely hampered given degraded utterances. To this end, we then propose speech enhancement concatenation and denoising training to improve the robustness. In addition to common degradations, we also consider adversarial noises, which alter the model output significantly yet are human-imperceptible. It was shown that both concatenations with off-the-shelf speech enhancement models and denoising training on voice conversion models could improve the robustness, while each of them had pros and cons.

This paper introduces S3PRL-VC, an open-source voice conversion (VC) framework based on the S3PRL toolkit. In the context of recognition-synthesis VC, self-supervised speech representation (S3R) is valuable in its potential to replace the expensive supervised representation adopted by state-of-the-art VC systems. Moreover, we claim that VC is a good probing task for S3R analysis. In this work, we provide a series of in-depth analyses by benchmarking on the two tasks in VCC2020, namely intra-/cross-lingual any-to-one (A2O) VC, as well as an any-to-any (A2A) setting. We also provide comparisons between not only different S3Rs but also top systems in VCC2020 with supervised representations. Systematic objective and subjective evaluation were conducted, and we show that S3R is comparable with VCC2020 top systems in the A2O setting in terms of similarity, and achieves state-of-the-art in S3R-based A2A VC. We believe the extensive analysis, as well as the toolkit itself, contribute to not only the S3R community but also the VC community. The codebase is now open-sourced1.

Unsupervised speech recognition (unsupervised ASR) aims to learn the ASR system with non-parallel speech and text corpus only. Wav2vec-U [1] has shown promising results in unsupervised ASR by self-supervised speech representations coupled with Generative Adversarial Network (GAN) training, but the robustness of the unsupervised ASR framework is unknown. In this work, we further analyze the training robustness of unsupervised ASR on the domain mismatch scenarios in which the domains of unpaired speech and text are different. Three domain mismatch scenarios include: (1) using speech and text from different datasets, (2) utilizing noisy/spontaneous speech, and (3) adjusting the amount of speech and text data. We also quantify the degree of the domain mismatch by calculating the JS-divergence of phoneme n-gram be-tween the transcription of speech and text. This metric correlates with the performance highly. Experimental results show that do-main mismatch leads to inferior performance, but a self-supervised model pre-trained on the targeted speech domain can extract better representation to alleviate the performance drop.

Self-supervised Speech Models (S3Ms) have been proven successful in many speech downstream tasks, like ASR. However, how pretraining data affects S3Ms’ downstream behavior remains an unexplored issue. In this paper, we study how pre-training data affects S3Ms by pre-training models on biased datasets targeting different factors of speech, including gender, content, and prosody, and evaluate these pre-trained S3Ms on selected downstream tasks in SUPERB Benchmark. Our experiments show that S3Ms have tolerance toward gender bias. Moreover, we find that the content of speech has little impact on the performance of S3Ms across downstream tasks, but S3Ms do show a preference toward a slower speech rate.

Automatic speaker verification (ASV), one of the most important technology for biometric identification, has been widely adopted in security-critical applications. However, ASV is seriously vulnerable to recently emerged adversarial attacks, yet effective counter-measures against them are limited. In this paper, we adopt neural vocoders to spot adversarial samples for ASV. We use the neural vocoder to re-synthesize audio and find that the difference between the ASV scores for the original and re-synthesized audio is a good indicator for discrimination between genuine and adversarial samples. This effort is, to the best of our knowledge, among the first to pursue such a technical direction for detecting time-domain adversarial samples for ASV, and hence there is a lack of established baselines for comparison. Consequently, we implement the Griffin-Lim algorithm as the detection baseline. The proposed approach achieves effective detection performance that outperforms the baselines in all the settings. We also show that the neural vocoder adopted in the detection framework is dataset-independent. Our codes will be made open-source for future works to do fair comparison1.

The past few years have witnessed the significant advances of speech synthesis and voice conversion technologies. However, such technologies can undermine the robustness of broadly implemented biometric identification models and can be harnessed by in-the-wild attackers for illegal uses. The ASVspoof challenge mainly focuses on synthesized audios by advanced speech synthesis and voice conversion models, and replay attacks. Recently, the first Audio Deep Synthesis Detection challenge (ADD 2022) extends the attack scenarios into more aspects. Also, ADD 2022 is the first challenge to propose the partially fake audio detection task. Such brand new attacks are dangerous and how to tackle such attacks remains an open question. Thus, we propose a novel framework by introducing the question-answering (fake span discovery) strategy with the self-attention mechanism to detect partially fake audios. The proposed fake span detection module tasks the anti-spoofing model to predict the start and end positions of the fake clip within the partially fake audio, address the model’s attention into discovering the fake spans rather than other shortcuts with less generalization, and finally equips the model with the discrimination capacity between real and partially fake audios. Our submission ranked second in the partially fake audio detection track of ADD 2022.

A leaderboard named Speech processing Universal PERformance Benchmark (SUPERB), which aims at benchmarking the performance of a shared self-supervised learning (SSL) speech model across various downstream speech tasks with minimal modification of architectures and a small amount of data, has fueled the research for speech representation learning. The SUPERB demonstrates speech SSL upstream models improve the performance of various downstream tasks through just minimal adaptation. As the paradigm of the self-supervised learning upstream model followed by downstream tasks arouses more attention in the speech community, characterizing the adversarial robustness of such paradigm is of high priority. In this paper, we make the first attempt to investigate the adversarial vulnerability of such paradigm under the attacks from both zero-knowledge adversaries and limited-knowledge adversaries. The experimental results illustrate that the paradigm proposed by SUPERB is seriously vulnerable to limited-knowledge adversaries, and the attacks generated by zero-knowledge adversaries are with transferability. The XAB test verifies the imperceptibility of crafted adversarial attacks.

Simultaneous speech translation (SimulST) is a challenging task aiming to translate streaming speech before the complete input is observed. A SimulST system generally includes two components: the pre-decision that aggregates the speech information and the policy that decides to read or write. While recent works had proposed various strategies to improve the pre-decision, they mainly adopt the fixed wait-k policy, leaving the adaptive policies rarely explored. This paper proposes to model the adaptive policy by adapting the Continuous Integrate-and-Fire (CIF). Compared with monotonic multihead attention (MMA), our method has the advantage of simpler computation, superior quality at low latency, and better generalization to long utterances. We conduct experiments on the MuST-C V2 dataset and show the effectiveness of our approach.

Speech representations learned from Self-supervised learning (SSL) models can benefit various speech processing tasks. However, utilizing SSL representations usually requires fine-tuning the pre-trained models or designing task-specific downstream models and loss functions, causing much memory usage and human labor. Recently, prompting in Natural Language Processing (NLP) has been found to be an efficient technique to leverage pre-trained language models (LMs). Specifically, prompt tuning optimizes a limited number of task-specific parameters with a fixed pre-trained model; as a result, only a small set of parameters is needed to be stored for each task. Prompt tuning improves computation and memory efficiency by leveraging the pre-trained LM's prediction ability. Nevertheless, such a paradigm is little studied in the speech community. We report in this paper the first exploration of the prompt tuning paradigm for speech processing tasks based on Generative Spoken Language Model (GSLM). Experiment results show that the prompt tuning technique achieves competitive performance in speech classification tasks with fewer trainable parameters than fine-tuning specialized downstream models. We further study the technique in challenging sequence generation tasks. Prompt tuning also demonstrates its potential, while the limitation and possible research directions are discussed in this paper. The source code is available on https://github.com/ga642381/SpeechPrompt.

This paper studies a transferable phoneme embedding framework that aims to deal with the cross-lingual text-to-speech (TTS) problem under the few-shot setting.Transfer learning is a common approach when it comes to few-shot learning since training from scratch on few-shot training data is bound to overfit.Still, we find that the naive transfer learning approach fails to adapt to unseen languages under extremely fewshot settings, where less than 8 minutes of data is provided.We deal with the problem by proposing a framework that consists of a phoneme-based TTS model and a codebook module to project phonemes from different languages into a learned latent space.Furthermore, by utilizing phoneme-level averaged selfsupervised learned features, we effectively improve the quality of synthesized speeches.Experiments show that using 4 utterances, which is about 30 seconds of data, is enough to synthesize intelligible speech when adapting to an unseen language using our framework.

Although deep learning-based end-to-end Automatic Speech Recognition (ASR) has shown remarkable performance in recent years, it suffers severe performance regression on test samples drawn from different data distributions.Test-time Adaptation (TTA), previously explored in the computer vision area, aims to adapt the model trained on source domains to yield better predictions for test samples, often out-of-domain, without accessing the source data.Here, we propose the Single-Utterance Test-time Adaptation (SUTA 1 ) framework for ASR, which is the first TTA study on ASR to our best knowledge.The single-utterance TTA is a more realistic setting that does not assume test data are sampled from identical distribution and does not delay on-demand inference due to pre-collection for the batch of adaptation data.SUTA consists of unsupervised objectives with an efficient adaptation strategy.Empirical results demonstrate that SUTA effectively improves the performance of the source ASR model evaluated on multiple out-of-domain target corpora and in-domain test samples.

Spoken Question Answering (SQA) is to find the answer from a spoken document given a question, which is crucial for personal assistants when replying to the queries from the users.Existing SQA methods all rely on Automatic Speech Recognition (ASR) transcripts.Not only does ASR need to be trained with massive annotated data that are time and cost-prohibitive to collect for low-resourced languages, but more importantly, very often the answers to the questions include name entities or out-of-vocabulary words that cannot be recognized correctly.Also, ASR aims to minimize recognition errors equally over all words, including many function words irrelevant to the SQA task.Therefore, SQA without ASR transcripts (textless) is always highly desired, although known to be very difficult.This work proposes Discrete Spoken Unit Adaptive Learning (DUAL), leveraging unlabeled data for pre-training and finetuned by the SQA downstream task.The time intervals of spoken answers can be directly predicted from spoken documents.We also release a new SQA benchmark corpus, NMSQA, for data with more realistic scenarios.We empirically showed that DUAL yields results comparable to those obtained by cascading ASR and text QA model and robust to real-world data.Our code and model will be open-sourced 1 .

Recently, adapting the idea of self-supervised learning (SSL) on continuous speech has started gaining attention.SSL models pre-trained on a huge amount of unlabeled audio can generate general-purpose representations that benefit a wide variety of speech processing tasks.Despite their ubiquitous deployment, however, the potential privacy risks of these models have not been well investigated.In this paper, we present the first privacy analysis on several SSL speech models using Membership Inference Attacks (MIA) under black-box access.The experiment results show that these pre-trained models are vulnerable to MIA and prone to membership information leakage with high Area Under the Curve (AUC) in both utterance-level and speaker-level.Furthermore, we also conduct several ablation studies to understand the factors that contribute to the success of MIA.

Recently, many novel techniques have been introduced to deal with spoofing attacks, and achieve promising countermeasure (CM) performances.However, these works only take the standalone CM models into account.Nowadays, a spoofing aware speaker verification (SASV) challenge which aims to facilitate the research of integrated CM and ASV models, arguing that jointly optimizing CM and ASV models will lead to better performance, is taking place.In this paper, we propose a novel multi-model and multi-level fusion strategy to tackle the SASV task.Compared with purely scoring fusion and embedding fusion methods, this framework first utilizes embeddings from CM models, propagating CM embeddings into a CM block to obtain a CM score.In the second-level fusion, the CM score and ASV scores directly from ASV systems will be concatenated into a prediction block for the final decision.As a result, the best single fusion system has achieved the SASV-EER of 0.97% on the evaluation set.Then by ensembling the top-5 fusion systems, the final SASV-EER reached 0.89%.

In this paper, we present Multi-scale Feature Aggregation Conformer (MFA-Conformer), an easy-to-implement, simple but effective backbone for automatic speaker verification based on the Convolution-augmented Transformer (Conformer).The architecture of the MFA-Conformer is inspired by recent stateof-the-art models in speech recognition and speaker verification.Firstly, we introduce a convolution subsampling layer to decrease the computational cost of the model.Secondly, we adopt Conformer blocks which combine Transformers and convolution neural networks (CNNs) to capture global and local features effectively.Finally, the output feature maps from all Conformer blocks are concatenated to aggregate multi-scale representations before final pooling.We evaluate the MFA-Conformer on the widely used benchmarks.The best system obtains 0.64%, 1.29% and 1.63% EER on VoxCeleb1-O, SITW.Dev, and SITW.Eval set, respectively.MFA-Conformer significantly outperforms the popular ECAPA-TDNN systems in both recognition performance and inference speed.Last but not the least, the ablation studies clearly demonstrate that the combination of global and local feature learning can lead to robust and accurate speaker embedding extraction.We have also released the code 1 for future comparison.

Deep learning has been the mainstream technique in natural language processing (NLP) area.However, the techniques require many labeled data and are less generalizable across domains.Meta-learning is an arising field in machine learning studying approaches to learn better learning algorithms.Approaches aim at improving algorithms in various aspects, including data efficiency and generalizability.Efficacy of approaches has been shown in many NLP tasks, but there is no systematic survey of these approaches in NLP, which hinders more researchers from joining the field.Our goal with this survey paper is to offer researchers pointers to relevant meta-learning works in NLP and attract more attention from the NLP community to drive future innovation.This paper first introduces the general concepts of meta-learning and the common approaches.Then we summarize task construction settings and application of meta-learning for various NLP problems and review the development of meta-learning in NLP community.

In this study, we aim to explore efficient tuning methods for speech self-supervised learning. Recent studies show that self-supervised learning (SSL) can learn powerful representations for different speech tasks. However, fine-tuning pre-trained models for each downstream task is parameter-inefficient since SSL models are notoriously large with millions of parameters. Adapters are lightweight modules commonly used in NLP to solve this problem. In downstream tasks, the parameters of SSL models are frozen, and only the adapters are trained. Given the lack of studies generally exploring the effectiveness of adapters for self-supervised speech tasks, we intend to fill this gap by adding various adapter modules in pre-trained speech SSL models. We show that the performance parity can be achieved with over 90% parameter reduction, and discussed the pros and cons of efficient tuning techniques. This is the first comprehensive investigation of various adapter types across speech tasks.

Audio-visual active speaker detection (AVASD) is well-developed, and now is an indispensable front-end for several multi-modal applications. However, to the best of our knowledge, the adversarial robustness of AVASD models hasn't been investigated, not to mention the effective defense against such attacks. In this paper, we are the first to reveal the vulnerability of AVASD models under audio-only, visual-only, and audio-visual adversarial attacks through extensive experiments. What's more, we also propose a novel audio-visual interaction loss (AVIL) for making attackers difficult to find feasible adversarial examples under an allocated attack budget. The loss aims at pushing the inter-class embeddings to be dispersed, namely non-speech and speech clusters, sufficiently disentangled, and pulling the intra-class embeddings as close as possible to keep them compact. Experimental results show the AVIL outperforms the adversarial training by 33.14 mAP (%) under multi-modal attacks.

We present the SUPERB challenge at SLT 2022, which aims at learning self-supervised speech representation for better performance, generalization, and efficiency. The challenge builds upon the SUPERB benchmark and implements metrics to measure the computation requirements of self-supervised learning (SSL) representation and to evaluate its generalizability and performance across the diverse SUPERB tasks. The SUPERB benchmark provides comprehensive coverage of popular speech processing tasks, from speech and speaker recognition to audio generation and semantic understanding. As SSL has gained interest in the speech community and showed promising outcomes, we envision the challenge to uplevel the impact of SSL techniques by motivating more practical designs of techniques beyond task performance. We summarize the results of 14 submitted models in this paper. We also discuss the main findings from those submissions and the future directions of SSL research.

Self-supervised learned (SSL) speech pre-trained models perform well across various speech processing tasks. Distilled versions of SSL models have been developed to match the needs of on-device speech applications. Though having similar performance as original SSL models, distilled counterparts suffer from performance degradation even more than their original versions in distorted environments. This paper proposes to apply Cross-Distortion Mapping and Domain Adversarial Training to SSL models during knowledge distillation to alleviate the performance gap caused by the domain mismatch problem. Results show consistent performance improvements under both in- and out-of-domain distorted setups for different downstream tasks while keeping efficient model size.

User-defined keyword spotting is a task to detect new spoken terms defined by users. This can be viewed as a few-shot learning problem since it is unreasonable for users to define their desired keywords by providing many examples. To solve this problem, previous works try to incorporate self-supervised learning models or apply meta-learning algorithms. But it is unclear whether self-supervised learning and meta-learning are complementary and which combination of the two types of approaches is most effective for few-shot keyword discovery. In this work, we systematically study these questions by utilizing various self-supervised learning models and combining them with a wide variety of meta-learning algorithms. Our result shows that HuBERT combined with Matching network achieves the best result and is robust to the changes of few-shot examples.

Self-Supervised Learning (SSL) from speech data has produced models that have achieved remarkable performance in many tasks, and that are known to implicitly represent many aspects of information latently present in speech signals. However, relatively little is known about the suitability of such models for prosody-related tasks or the extent to which they encode prosodic information. We present a new evaluation framework, “SUPERB-prosody,” consisting of three prosody-related downstream tasks and two pseudo tasks. We find that 13 of the 15 SSL models outperformed the baseline on all the prosody-related tasks. We also show good performance on two pseudo tasks: prosody reconstruction and future prosody prediction. We further analyze the layerwise contributions of the SSL models. Overall we conclude that SSL speech models are highly effective for prosody-related tasks. We release our code11https://github.com/JSALT-2022-SSL/superb-prosody for the community to support further investigation of SSL models' utility for prosody.

Compressing self-supervised models has become increasingly necessary, as self-supervised models become larger. While previous approaches have primarily focused on compressing the model size, shortening sequences is also effective in reducing the computational cost. In this work, we study fixed-length and variable-length subsampling along the time axis in self-supervised learning. We explore how individual downstream tasks are sensitive to input frame rates. Subsampling while training self-supervised models not only improves the overall performance on downstream tasks under certain frame rates, but also brings significant speed-up in inference. Variable-length subsampling performs particularly well under low frame rates. In addition, if we have access to phonetic boundaries, we find no degradation in performance for an average frame rate as low as 10 Hz.

Data-driven speech processing models usually perform well with a large amount of text supervision, but collecting transcribed speech data is costly. Therefore, we propose Speech-CLIP, a novel framework bridging speech and text through images to enhance speech models without transcriptions. We leverage state-of-the-art pre-trained HuBERT and CLIP, aligning them via paired images and spoken captions with minimal fine-tuning. SpeechCLIP outperforms prior state-of-the-art on image-speech retrieval and performs zero-shot speech-text retrieval without direct supervision from transcriptions. Moreover, SpeechCLIP can directly retrieve semantically related keywords from speech.

Personalizing a speech synthesis system is a highly desired application, where the system can generate speech with the user’s voice with rare enrolled recordings. There are two main approaches to build such a system in recent works: speaker adaptation and speaker encoding. On the one hand, speaker adaptation methods fine-tune a trained multi-speaker text-to-speech (TTS) model with few enrolled samples. However, they require at least thousands of fine-tuning steps for high-quality adaptation, making it hard to apply on devices. On the other hand, speaker encoding methods encode enrollment utterances into a speaker embedding. The trained TTS model can synthesize the user’s speech conditioned on the corresponding speaker embedding. Nevertheless, the speaker encoder suffers from the generalization gap between the seen and unseen speakers. In this paper, we propose applying a meta-learning algorithm to the speaker adaptation method. More specifically, we use Model Agnostic Meta-Learning (MAML) as the training algorithm of a multi-speaker TTS model, which aims to find a great meta-initialization to adapt the model to any few-shot speaker adaptation tasks quickly. Therefore, we can also adapt the meta-trained TTS model to unseen speakers efficiently. Our experiments compare the proposed method (Meta-TTS) with two baselines: a speaker adaptation method baseline and a speaker encoding method baseline. The evaluation results show that Meta-TTS can synthesize high speaker-similarity speech from few enrollment samples with fewer adaptation steps than the speaker adaptation baseline and outperforms the speaker encoding baseline under the same training scheme. When the speaker encoder of the baseline is pre-trained with extra 8371 speakers of data, Meta-TTS can still outperform the baseline on LibriTTS dataset and achieve comparable results on VCTK dataset.

ASRhas been shown to achieve great performance recently. However, most of them rely on massive paired data, which is not feasible for low-resource languages worldwide. This paper investigates how to learn directly from unpaired phone sequences and speech utterances. We design a two-stage iterative framework. GAN training is adopted in the first stage to find the mapping relationship between unpaired speech and phone sequence. In the second stage, another HMM model is introduced to train from the generator’s output, which boosts the performance and provides a better segmentation for the next iteration. In the experiment, we first investigate different choices of model designs. Thenwe compare the framework to different types of baselines: (i) supervised methods (ii) acoustic unit discovery based methods (iii) methods learning from unpaired data. Our framework performs consistently better than all acoustic unit discovery methods and previous methods learning from unpaired data based on the TIMIT dataset.

Previous works have shown that automatic speaker verification (ASV) is seriously vulnerable to malicious spoofing attacks, such as replay, synthetic speech, and recently emerged adversarial attacks. Great efforts have been dedicated to defending ASV against replay and synthetic speech; however, only a few approaches have been explored to deal with adversarial attacks. All the existing approaches to tackle adversarial attacks for ASV require the knowledge for adversarial samples generation, but it is impractical for defenders to know the exact attack algorithms that are applied by the in-the-wild attackers. This work is among the first to perform adversarial defense for ASV without knowing the specific attack algorithms. Inspired by self-supervised learning models (SSLMs) that possess the merits of alleviating the superficial noise in the inputs and reconstructing clean samples from the interrupted ones, this work regards adversarial perturbations as one kind of noise and conducts adversarial defense for ASV by SSLMs. Specifically, we propose to perform adversarial defense from two perspectives: 1) adversarial perturbation purification and 2) adversarial perturbation detection. The purification module aims at alleviating the adversarial perturbations in the samples and pulling the contaminated adversarial inputs back towards the decision boundary. Experimental results show that our proposed purification module effectively counters adversarial attacks and outperforms traditional filters from both alleviating the adversarial noise and maintaining the performance of genuine samples. The detection module aims at detecting adversarial samples from genuine ones based on the statistical properties of ASV scores derived by a unique ASV integrating with different number of SSLMs. Experimental results show that our detection module helps shield the ASV by detecting adversarial samples. Both purification and detection methods are helpful for defending against different kinds of attack algorithms. Moreover, since there is no common metric for evaluating the ASV performance under adversarial attacks, this work also formalizes evaluation metrics for adversarial defense considering both purification and detection based approaches into account. We sincerely encourage future works to benchmark their approaches based on the proposed evaluation framework.

Question-answering systems are becoming increasingly popular in Natural Language Processing, especially when applied in smart factory settings. A common practice in designing those systems is through intent classification. However, in a multiple-stage task commonly seen in those settings, relying solely on intent classification may lead to erroneous answers, as questions rising from different work stages may share the same intent but have different contexts and therefore require different answers. To address this problem, we designed an interactive dialogue system that utilizes contextual information to assist intent classification in a multiple-stage task. Specifically, our system incorporates user’s utterances with real-time video feed to better situate users’ questions and analyze their intent.

Self-supervised pretraining on speech data has achieved a lot of progress. High-fidelity representation of the speech signal is learned from a lot of untranscribed data and shows promising performance. Recently, there are several works focusing on evaluating the quality of self-supervised pretrained representations on various tasks with-out domain restriction, e.g. SUPERB. However, such evaluations do not provide a comprehensive comparison among many ASR benchmark corpora. In this paper, we focus on the general applications of pretrained speech representations, on advanced end-to-end automatic speech recognition (E2E-ASR) models. We select sev-eral pretrained speech representations and present the experimental results on various open-source and publicly available corpora for E2E-ASR. Without any modification of the back-end model archi-tectures or training strategy, some of the experiments with pretrained representations, e.g., WSJ, WSJ0-2mix with HuBERT, reach or out-perform current state-of-the-art (SOTA) recognition performance. Moreover, we further explore more scenarios for whether the pre-training representations are effective, such as the cross-language or overlapped speech. The scripts, configuratons and the trained mod-els have been released in ESPnet to let the community reproduce our experiments and improve them.

Mandarin-English code-switching (CS) is frequently used among East and Southeast Asian people. However, the intra-sentence language switching of the two very different languages makes recognizing CS speech challenging. Meanwhile, the recent successful non-autoregressive (NAR) ASR models remove the need for left-to-right beam decoding in autoregressive (AR) models and achieved outstanding performance and fast inference speed, but it has not been applied to Mandarin-English CS speech recognition. This paper takes advantage of the Mask-CTC NAR ASR framework to tackle the CS speech recognition issue. We further propose to change the Mandarin output target of the encoder to Pinyin for faster encoder training and introduce the Pinyin-to-Mandarin decoder to learn contextualized information. Moreover, we use word embedding label smoothing to regularize the decoder with contextualized information and projection matrix regularization to bridge that gap between the encoder and decoder. We evaluate these methods on the SEAME corpus and achieved exciting results.

Recently, voice conversion (VC) has been widely studied. Many VC systems use disentangle-based learning techniques to separate the speaker and the linguistic content information from a speech signal. Subsequently, they convert the voice by changing the speaker information to that of the target speaker. To prevent the speaker information from leaking into the content embeddings, previous works either reduce the dimension or quantize the content embedding as a strong information bottleneck. These mechanisms somehow hurt the synthesis quality. In this work, we propose AGAIN-VC, an innovative VC system using Activation Guidance and Adaptive Instance Normalization. AGAIN-VC is an auto-encoder-based model, comprising of a single encoder and a decoder. With a proper activation as an information bottleneck on content embeddings, the trade-off between the synthesis quality and the speaker similarity of the converted speech is improved drastically. This one-shot VC system obtains the best performance regardless of the subjective or objective evaluations.

The few-shot multi-speaker multi-style voice cloning task is to synthesize utterances with voice and speaking style similar to a reference speaker given only a few reference samples. In this work, we investigate different speaker representations and proposed to integrate pretrained and learnable speaker representations. Among different types of embeddings, the embedding pretrained by voice conversion achieves the best performance. The FastSpeech 2 model combined with both pretrained and learnable speaker representations shows great generalization ability on few-shot speakers and achieved 2nd place in the one-shot track of the ICASSP 2021 M2VoC challenge.

Much recent work on Spoken Language Understanding (SLU) is limited in at least one of three ways: models were trained on oracle text input and neglected ASR errors, models were trained to predict only intents without the slot values, or models were trained on a large amount of in-house data. In this paper, we propose a clean and general framework to learn semantics directly from speech with semi-supervision from transcribed or untranscribed speech to address these issues. Our framework is built upon pretrained end-to-end (E2E) ASR and self-supervised language models, such as BERT, and fine-tuned on a limited amount of target SLU data. We study two semi-supervised settings for the ASR component: supervised pretraining on transcribed speech, and unsupervised pretraining by replacing the ASR encoder with self-supervised speech representations, such as wav2vec. In parallel, we identify two essential criteria for evaluating SLU models: environmental noise-robustness and E2E semantics evaluation. Experiments on ATIS show that our SLU framework with speech as input can perform on par with those using oracle text as input in semantics understanding, even though environmental noise is present and a limited amount of labeled semantics data is available for training.

Any-to-any voice conversion aims to convert the voice from and to any speakers even unseen during training, which is much more challenging compared to one-to-one or many-to-many tasks, but much more attractive in real-world scenarios. In this paper we proposed FragmentVC, in which the latent phonetic structure of the utterance from the source speaker is obtained from Wav2Vec 2.0, while the spectral features of the utterance(s) from the target speaker are obtained from log mel-spectrograms. By aligning the hidden structures of the two different feature spaces with a two-stage training process, FragmentVC is able to extract fine-grained voice fragments from the target speaker utterance(s) and fuse them into the desired utterance, all based on the attention mechanism of Transformer as verified with analysis on attention maps, and is accomplished end-to-end. This approach is trained with reconstruction loss only without any disentanglement considerations between content and speaker information and doesn't require parallel data. Objective evaluation based on speaker verification and subjective evaluation with MOS both showed that this approach outperformed SOTA approaches, such as AdaIN-VC and AutoVC.

Despite the recent success of speech separation models, they fail to separate sources properly while facing different sets of people or noisy environments. To tackle this problem, we proposed to apply meta-learning to the speech separation task. We aimed to find a meta-initialization model, which can quickly adapt to new speakers by seeing only one mixture generated by those people. In this paper, we use model-agnostic meta-learning(MAML) algorithm and almost no inner loop(ANIL) algorithm in Conv-TasNet to achieve this goal. The experiment results show that our model can adapt not only to a new set of speakers but also noisy environments. Furthermore, we found out that the encoder and decoder serve as the feature-reuse layers, while the separator is the task-specific module.

Automatic speaker verification (ASV) is one of the core technologies in biometric identification. With the ubiquitous usage of ASV systems in safety-critical applications, more and more malicious attackers attempt to launch adversarial attacks at ASV systems. In the midst of the arms race between attack and defense in ASV, how to effectively improve the robustness of ASV against adversarial attacks remains an open question. We note that the self-supervised learning models possess the ability to mitigate superficial perturbations in the input after pretraining. Hence, with the goal of effective defense in ASV against adversarial attacks, we propose a standard and attack-agnostic method based on cascaded self-supervised learning models to purify the adversarial perturbations. Experimental results demonstrate that the proposed method achieves effective defense performance and can successfully counter adversarial attacks in scenarios where attackers may either be aware or unaware of the self-supervised learning models.

Automatic speech recognition (ASR) technologies today are primarily optimized for given datasets; thus, any changes in the application environment (e.g., acoustic conditions or topic domains) may inevitably degrade the performance. We can collect new data describing the new environment and fine-tune the system, but this naturally leads to higher error rates for the earlier datasets, referred to as catastrophic forgetting. The concept of lifelong learning (LLL) aiming to enable a machine to sequentially learn new tasks from new datasets describing the changing real world without forgetting the previously learned knowledge is thus brought to attention. This paper reports, to our knowledge, the first effort to extensively consider and analyze the use of various approaches of LLL in end-to-end (E2E) ASR, including proposing novel methods in saving data for past domains to mitigate the catastrophic forgetting problem. An overall relative reduction of 28.7% in WER was achieved compared to the fine-tuning baseline when sequentially learning on three very different benchmark corpora. This can be the first step toward the highly desired ASR technologies capable of synchronizing with the continuously changing real world.

Speech separation has been well developed, with the very successful permutation invariant training (PIT) approach, although the frequent label assignment switching happening during PIT training remains to be a problem when better convergence speed and achievable performance are desired.In this paper, we propose to perform self-supervised pre-training to stabilize the label assignment in training the speech separation model.Experiments over several types of self-supervised approaches, several typical speech separation models and two different datasets showed that very good improvements are achievable if a proper self-supervised approach is chosen.

Any-to-any voice conversion (VC) aims to convert the timbre of utterances from and to any speakers seen or unseen during training.Various any-to-any VC approaches have been proposed like AUTOVC, AdaINVC, and FragmentVC.AUTOVC, and AdaINVC utilize source and target encoders to disentangle the content and speaker information of the features.Frag-mentVC utilizes two encoders to encode source and target information and adopts cross attention to align the source and target features with similar phonetic content.Moreover, pretrained features are adopted.AUTOVC used d-vector to extract speaker information, and self-supervised learning (SSL) features like wav2vec 2.0 is used in FragmentVC to extract the phonetic content information.Different from previous works, we proposed S2VC that utilizes Self-Supervised features as both source and target features for the VC model.Supervised phoneme posteriorgram (PPG), which is believed to be speaker-independent and widely used in VC to extract content information, is chosen as a strong baseline for SSL features.The objective evaluation and subjective evaluation both show models taking SSL feature CPC as both source and target features outperforms that taking PPG as source feature, suggesting that SSL features have great potential in improving VC.

Speech quality assessment has been a critical issue in speech processing for decades. Existing automatic evaluations usually require clean references or parallel ground truth data, which is infeasible when the amount of data soars. Subjective tests, on the other hand, do not need any additional clean or parallel data and correlates better to human perception. However, such a test is expensive and time-consuming because crowd work is necessary. It thus becomes highly desired to develop an automatic evaluation approach that correlates well with human perception while not requiring ground truth data. In this paper, we use self-supervised pre-trained models for MOS prediction. We show their representations can distinguish between clean and noisy audios. Then, we fine-tune these pre-trained models followed by simple linear layers in an end-to-end manner. The experiment results showed that our framework outperforms the two previous state-of-the-art models by a significant improvement on Voice Conversion Challenge 2018 and achieves comparable or superior performance on Voice Conversion Challenge 2016. We also conducted an ablation study to further investigate how each module benefits the task. The experiment results are implemented and reproducible with publicly available toolkits.

Auto-KWS 2021 challenge calls for automated machine learning (AutoML) solutions to automate the process of applying machine learning to a customized keyword spotting task.Compared with other keyword spotting tasks, Auto-KWS challenge has the following three characteristics: 1) The challenge focuses on the problem of customized keyword spotting, where the target device can only be awakened by an enrolled speaker with his specified keyword.The speaker can use any language and accent to define his keyword.2) All dataset of the challenge is recorded in realistic environment.It is to simulate different user scenarios.3) Auto-KWS is a "code competition", where participants need to submit AutoML solutions, then the platform automatically runs the enrollment and prediction steps with the submitted code.This challenge aims at promoting the development of a more personalized and flexible keyword spotting system.Two baseline systems are provided to all participants as references.

Automatic speaker verification (ASV) is a well developed technology for biometric identification, and has been ubiquitous implemented in security-critic applications, such as banking and access control.However, previous works have shown that ASV is under the radar of adversarial attacks, which are very similar to their original counterparts from human's perception, yet will manipulate the ASV render wrong prediction.Due to the very late emergence of adversarial attacks for ASV, effective countermeasures against them are limited.Given that the security of ASV is of high priority, in this work, we propose the idea of "voting for the right answer" to prevent risky decisions of ASV in blind spot areas, by employing random sampling and voting.Experimental results show that our proposed method improves the robustness against both the limited-knowledge attackers by pulling the adversarial samples out of the blind spots, and the sufficient-knowledge attackers by introducing randomness and increasing the attackers' budgets.

Self-supervised learning (SSL) has proven vital for advancing research in natural language processing (NLP) and computer vision (CV).The paradigm pretrains a shared model on large volumes of unlabeled data and achieves state-of-the-art (SOTA) for various tasks with minimal adaptation.However, the speech processing community lacks a similar setup to systematically explore the paradigm.To bridge this gap, we introduce Speech processing Universal PERformance Benchmark (SUPERB).SUPERB is a leaderboard to benchmark the performance of a shared model across a wide range of speech processing tasks with minimal architecture changes and labeled data.Among multiple usages of the shared model, we especially focus on extracting the representation learned from SSL for its preferable re-usability.We present a simple framework to solve SUPERB tasks by learning task-specialized lightweight prediction heads on top of the frozen shared model.Our results demonstrate that the framework is promising as SSL representations show competitive generalizability and accessibility across SUPERB tasks.We release SUPERB as a challenge with a leaderboard 1 and a benchmark toolkit 2 to fuel the research in representation learning and general speech processing.

Hsuan Su, Jiun-Hao Jhan, Fan-yun Sun, Saurav Sahay, Hung-yi Lee. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Whispering is an important mode of human speech, but no end-to-end recognition results for it were reported yet, probably due to the scarcity of available whispered speech data. In this paper, we present several approaches for end-to-end (E2E) recognition of whispered speech considering the special characteristics of whispered speech and the scarcity of data. This includes a frequency-weighted SpecAugment policy and a frequency-divided CNN feature extractor for better capturing the high-frequency structures of whispered speech, and a layer-wise transfer learning approach to pre-train a model with normal or normal-to-whispered converted speech then fine-tune it with whispered speech to bridge the gap between whispered and normal speech. We achieve an overall relative reduction of 19.8% in PER and 44.4% in CER on a relatively small whispered TIMIT corpus. The results indicate as long as we have a good E2E model pre-trained on normal or pseudo-whispered speech, a relatively small set of whispered speech may suffice to obtain a reasonably good E2E whispered speech recognizer.

Self-supervised speech models are powerful speech representation extractors for downstream applications. Recently, larger models have been utilized in acoustic model training to achieve better performance. We propose Audio ALBERT, a lite version of the self-supervised speech representation model. We apply the lightweight representation extractor to two downstream tasks, speaker classification and phoneme classification. We show that Audio ALBERT achieves performance comparable with massive pre-trained networks in the downstream tasks while having 91% fewer parameters. Moreover, we design probing models to measure how much the latent representations can encode the speaker's and phoneme's information. We find that the representations encoded in internal layers of Audio ALBERT contain more information for both phoneme and speaker than the last layer, which is generally used for downstream tasks. Our findings provide a new avenue for using self-supervised networks to achieve better performance and efficiency.

Prosody modeling is an essential component in modern text-to-speech (TTS) frameworks. By explicitly providing prosody features to the TTS model, the style of synthesized utterances can thus be controlled. However, predicting natural and reasonable prosody at inference time is challenging. In this work, we analyzed the behavior of non-autoregressive TTS models under different prosody-modeling settings and proposed a hierarchical architecture, in which the prediction of phoneme-level prosody features are conditioned on the word-level prosody features. The proposed method outperforms other competitors in terms of audio quality and prosody naturalness on objective and subjective evaluation.

Voice conversion technologies have been greatly improved in recent years with the help of deep learning, but their capabilities of producing natural sounding utterances in different conditions remain unclear. In this paper, we gave a thorough study of the robustness of known VC models. We also modified these models, such as the replacement of speaker embeddings, to further improve their performances. We found that the sampling rate and audio duration greatly influence voice conversion. All the VC models suffer from unseen data, but AdaIN-VC is relatively more robust. Also, the speaker embedding jointly trained is more suitable for voice conversion than those trained on speaker identification.

Substantial improvements have been achieved in recent years in voice conversion, which converts the speaker characteristics of an utterance into those of another speaker without changing the linguistic content of the utterance. Nonetheless, the improved conversion technologies also led to concerns about privacy and authentication. It thus becomes highly desired to be able to prevent one's voice from being improperly utilized with such voice conversion technologies. This is why we report in this paper the first known attempt to perform adversarial attack on voice conversion. We introduce human imperceptible noise into the utterances of a speaker whose voice is to be defended. Given these adversarial examples, voice conversion models cannot convert other utterances so as to sound like being produced by the defended speaker. Preliminary experiments were conducted on two currently state-of-the-art zero-shot voice conversion models. Objective and subjective evaluation results in both white-box and black-box scenarios are reported. It was shown that the speaker characteristics of the converted utterances were made obviously different from those of the defended speaker, while the adversarial examples of the defended speaker are not distinguishable from the authentic utterances.

In this article, we target speech translation (ST). We propose lightweight approaches that generally improve either ASR or end-to-end ST models. We leverage continuous representations of words, known as word embeddings, to improve ASR in cascaded systems as well as end-to-end ST models. The benefit of using word embedding is that word embedding can be obtained easily by training on pure textual data, which alleviates data scarcity issue. Also, word embedding provides additional contextual information to speech models. We motivate to distill the knowledge from word embedding into speech models. In ASR, we use word embeddings as a regularizer to reduce the WER, and further propose a novel decoding method to fuse the semantic relations among words for further improvement. In the end-to-end ST model, we propose leveraging word embeddings as an intermediate representation to enhance translation performance. Our analysis shows that it is possible to map speech signals to semantic space, which motivates future work on applying the proposed methods in spoken language processing tasks.

We introduce a self-supervised speech pre-training method called TERA, which stands for Transformer Encoder Representations from Alteration. Recent approaches often learn by using a single auxiliary task like contrastive prediction, autoregressive prediction, or masked reconstruction. Unlike previous methods, we use alteration along three orthogonal axes to pre-train Transformer Encoders on a large amount of unlabeled speech. The model learns through the reconstruction of acoustic frames from their altered counterpart, where we use a stochastic policy to alter along various dimensions: time, frequency, and magnitude. TERA can be used for speech representations extraction or fine-tuning with downstream models. We evaluate TERA on several downstream tasks, including phoneme classification, keyword spotting, speaker recognition, and speech recognition. We present a large-scale comparison of various self-supervised models. TERA achieves strong performance in the comparison by improving upon surface features and outperforming previous models. In our experiments, we study the effect of applying different alteration techniques, pre-training on more data, and pre-training on various features. We analyze different model sizes and find that smaller models are strong representation learners than larger models, while larger models are more effective for downstream fine-tuning than smaller models. Furthermore, we show the proposed method is transferable to downstream datasets not used in pre-training.

Multi-label classification (MLC) assigns multiple labels to each sample. Prior studies show that MLC can be transformed to a sequence prediction problem with a recurrent neural network (RNN) decoder to model the label dependency. However, training a RNN decoder requires a predefined order of labels, which is not directly available in the MLC specification. Besides, RNN thus trained tends to overfit the label combinations in the training set and have difficulty generating unseen label sequences. In this paper, we propose a new framework for MLC which does not rely on a predefined label order and thus alleviates exposure bias. The experimental results on three multi-label classification benchmark datasets show that our method outperforms competitive baselines by a large margin. We also find the proposed approach has a higher probability of generating label combinations not seen during training than the baseline models. The result shows that the proposed approach has better generalization capability.

Speech translation (ST) aims to learn transformations from speech in the source language to the text in the target language.Previous works show that multitask learning improves the ST performance, in which the recognition decoder generates the text of the source language, and the translation decoder obtains the final translations based on the output of the recognition decoder.Because whether the output of the recognition decoder has the correct semantics is more critical than its accuracy, we propose to improve the multitask ST model by utilizing word embedding as the intermediate.

While behaviors of pretrained language models (LMs) have been thoroughly examined, what happened during pretraining is rarely studied.We thus investigate the developmental process from a set of randomly initialized parameters to a totipotent 1 language model, which we refer to as the embryology of a pretrained language model.Our results show that ALBERT learns to reconstruct and predict tokens of different parts of speech (POS) in different learning speeds during pretraining.We also find that linguistic knowledge and world knowledge do not generally improve as pretraining proceeds, nor do downstream tasks' performance.These findings suggest that knowledge of a pretrained model varies during pretraining, and having more pretrain steps does not necessarily provide a model with more comprehensive knowledge.We provide source codes and pretrained models to reproduce our results at https://github. com/d223302/albert-embryology.

To rectify fisheye distortion from a single image, we advance self-supervised learning strategies and propose a unique deep learning model of Fisheye GAN (FE-GAN). Our FE-GAN learns pixel-level distortion flow from sets of fisheye distorted images and distortion-free ones (but not requiring such correspondences), with unique cross-rotation and intra-warping consistency introduced. With such novel self-supervised learning techniques, our FEGAN is able to recover the distortion-free image directly from the single fisheye image input. Our experiments quantitatively and qualitative confirm the effectiveness and robustness of our proposed model, which performs favorably against recent GAN-based image translation models.

Lack of code-switching data is an issue of training codeswitching language model. In this paper, we propose an approach to train code-switching language models with monolingual data only. By constraining and normalizing output projection matrix in RNN based language model, we make the embeddings of different languages close to each other. With the numerical and visualized results, we show that the proposed approaches remarkably improve the code-switching language modeling trained from monolingual data. The proposed approaches are comparable or even better than training code-switching language model with artificially generated code-switching data. Furthermore, we use unsupervised bilingual word translation to analyze if semantically equivalent words in different languages are mapped together.

In this paper, we proposed to apply meta learning approach for low-resource automatic speech recognition (ASR). We formulated ASR for different languages as different tasks, and meta-learned the initialization parameters from many pretraining languages to achieve fast adaptation on unseen target language, via recently proposed model-agnostic meta learning algorithm (MAML). We evaluated the proposed approach using six languages as pretraining tasks and four languages as target tasks. Preliminary results showed that the proposed method, MetaASR, significantly outperforms the state-of-the-art multitask pretraining approach on all target languages with different combinations of pretraining languages. In addition, since MAML's model-agnostic property, this paper also opens new research direction of applying meta learning to more speech-related applications.

End-to-end speech recognition systems have achieved competitive results compared to traditional systems. However, the complex transformations involved between layers given highly variable acoustic signals are hard to analyze. In this paper, we present our ASR probing model, which synthesizes speech from hidden representations of end-to-end ASR to examine the information maintained after each layer calculation. Listening to the synthesized speech, we observe gradual removal of speaker variability and noise as the layer goes deeper, which aligns with the previous studies on how deep network functions in speech recognition. This paper is the first study analyzing the end-to-end speech recognition model by demonstrating what each layer hears. Speaker verification and speech enhancement measurements on synthesized speech are also conducted to confirm our observation further.

In this paper, we investigate the benefit that off-the-shelf word embedding can bring to the sequence-to-sequence (seq-to-seq) automatic speech recognition (ASR). We first introduced the word embedding regularization by maximizing the cosine similarity between a transformed decoder feature and the target word embedding. Based on the regularized decoder, we further proposed the fused decoding mechanism. This allows the decoder to consider the semantic consistency during decoding by absorbing the information carried by the transformed decoder feature, which is learned to be close to the target word embedding. Initial results on LibriSpeech demonstrated that pre-trained word embedding can significantly lower ASR recognition error with a negligible cost, and the choice of word embedding algorithms among Skip-gram, CBOW and BERT is important.

In this paper we propose a Sequential Representation Quantization AutoEncoder (SeqRQ-AE) to learn from primarily unpaired audio data and produce sequences of representations very close to phoneme sequences of speech utterances. This is achieved by proper temporal segmentation to make the representations phoneme-synchronized, and proper phonetic clustering to have total number of distinct representations close to the number of phonemes. Mapping between the distinct representations and phonemes is learned from a small amount of annotated paired data. Preliminary experiments on LJSpeech demonstrated the learned representations for vowels have relative locations in latent space in good parallel to that shown in the IPA vowel chart defined by linguistics experts. With less than 20 minutes of annotated speech, our method outperformed existing methods on phoneme recognition and is able to synthesize intelligible speech that beats our baseline model.

We present Mockingjay as a new speech representation learning approach, where bidirectional Transformer encoders are pre-trained on a large amount of unlabeled speech. Previous speech representation methods learn through conditioning on past frames and predicting information about future frames. Whereas Mockingjay is designed to predict the current frame through jointly conditioning on both past and future contexts. The Mockingjay representation improves performance for a wide range of downstream tasks, including phoneme classification, speaker recognition, and sentiment classification on spoken content, while outperforming other approaches. Mockingjay is empirically powerful and can be fine-tuned with downstream models, with only 2 epochs we further improve performance dramatically. In a low resource setting with only 0.1% of labeled data, we outperform the result of Mel-features that uses all 100% labeled data.

In this paper, we propose a vector quantization (VQ) based one-shot voice conversion (VC) approach without any supervision on speaker label. We model the content embedding as a series of discrete codes and take the difference between quantize-before and quantize-after vector as the speaker embedding. We show that this approach has a strong ability to disentangle the content and speaker information with reconstruction loss only, and one-shot VC is thus achieved.

Various forefront countermeasure methods for automatic speaker verification (ASV) with considerable performance in anti-spoofing are proposed in the ASVspoof 2019 challenge. However, previous work has shown that countermeasure models are vulnerable to adversarial examples indistinguishable from natural data. A good countermeasure model should not only be robust against spoofing audio, including synthetic, converted, and replayed audios; but counteract deliberately generated examples by malicious adversaries. In this work, we introduce a passive defense method, spatial smoothing, and a proactive defense method, adversarial training, to mitigate the vulnerability of ASV spoofing countermeasure models against adversarial examples. This paper is among the first to use defense methods to improve the robustness of ASV spoofing countermeasure models under adversarial attacks. The experimental results show that these two defense methods positively help spoofing countermeasure models counter adversarial examples.

Permutation Invariant Training (PIT) has long been a stepping stone method for training speech separation model in handling the label ambiguity problem. With PIT selecting the minimum cost label assignments dynamically, very few studies considered the separation problem to be optimizing both the model parameters and the label assignments, but focused on searching for good model architecture and parameters. In this paper, we investigate instead for a given model architecture the various flexible label assignment strategies for training the model, rather than directly using PIT. Surprisingly, we discover a significant performance boost compared to PIT is possible if the model is trained with fixed label assignments and a good set of labels is chosen. With fixed label training cascaded between two sections of PIT, we achieved the state-of-the-art performance on WSJ0-2mix without changing the model architecture at all.

In previous works, only parameter weights of ASR models are optimized under fixed-topology architecture.However, the design of successful model architecture has always relied on human experience and intuition.Besides, many hyperparameters related to model architecture need to be manually tuned.Therefore in this paper, we propose an ASR approach with efficient gradient-based architecture search, DARTS-ASR.In order to examine the generalizability of DARTS-ASR, we apply our approach not only on many languages to perform monolingual ASR, but also on a multilingual ASR setting.Following previous works, we conducted experiments on a multilingual dataset, IARPA BABEL.The experiment results show that our approach outperformed the baseline fixed-topology architecture by 10.2% and 10.0% relative reduction on character error rates under monolingual and multilingual ASR settings respectively.Furthermore, we perform some analysis on the searched architectures by DARTS-ASR.

While various end-to-end models for spoken language understanding tasks have been explored recently, this paper is probably the first known attempt to challenge the very difficult task of end-to-end spoken question answering (SQA).Learning from the very successful BERT model for various text processing tasks, here we proposed an audio-and-text jointly learned SpeechBERT model.This model outperformed the conventional approach of cascading ASR with the following text question answering (TQA) model on datasets including ASR errors in answer spans, because the end-to-end model was shown to be able to extract information out of audio data before ASR produced errors.When ensembling the proposed end-to-end model with the cascade architecture, even better performance was achieved.In addition to the potential of end-to-end SQA, the SpeechBERT can also be considered for many other spoken language understanding tasks just as BERT for many text processing tasks.

In this paper, we propose WG-WaveNet, a fast, lightweight, and high-quality waveform generation model. WG-WaveNet is composed of a compact flow-based model and a post-filter. The two components are jointly trained by maximizing the likelihood of the training data and optimizing loss functions on the frequency domains. As we design a flow-based model that is heavily compressed, the proposed model requires much less computational resources compared to other waveform generation models during both training and inference time; even though the model is highly compressed, the post-filter maintains the quality of generated waveform. Our PyTorch implementation can be trained using less than 8 GB GPU memory and generates audio samples at a rate of more than 960 kHz on an NVIDIA 1080Ti GPU. Furthermore, even if synthesizing on a CPU, we show that the proposed method is capable of generating 44.1 kHz speech waveform 1.2 times faster than real-time. Experiments also show that the quality of generated audio is comparable to those of other methods. Audio samples are publicly available online.

Recently, end-to-end multi-speaker text-to-speech (TTS) systems gain success in the situation where a lot of high-quality speech plus their corresponding transcriptions are available.However, laborious paired data collection processes prevent many institutes from building multi-speaker TTS systems of great performance.In this work, we propose a semi-supervised learning approach for multi-speaker TTS.A multi-speaker TTS model can learn from the untranscribed audio via the proposed encoder-decoder framework with discrete speech representation.The experiment results demonstrate that with only an hour of paired speech data, whether the paired data is from multiple speakers or a single speaker, the proposed model can generate intelligible speech in different voices.We found the model can benefit from the proposed semi-supervised learning approach even when part of the unpaired speech data is noisy.In addition, our analysis reveals that different speaker characteristics of the paired data have an impact on the effectiveness of semisupervised TTS.

Voice conversion (VC) is a task that transforms the source speaker's timbre, accent, and tones in audio into another one's while preserving the linguistic content.It is still a challenging work, especially in a one-shot setting.Auto-encoder-based VC methods disentangle the speaker and the content in input speech without given the speaker's identity, so these methods can further generalize to unseen speakers.The disentangle capability is achieved by vector quantization (VQ), adversarial training, or instance normalization (IN).However, the imperfect disentanglement may harm the quality of output speech.In this work, to further improve audio quality, we use the U-Net architecture within an auto-encoder-based VC system.We find that to leverage the U-Net architecture, a strong information bottleneck is necessary.The VQ-based method, which quantizes the latent vectors, can serve the purpose.The objective and the subjective evaluations show that the proposed method performs well in both audio naturalness and speaker similarity.

High-performance anti-spoofing models for automatic speaker verification (ASV), have been widely used to protect ASV by identifying and filtering spoofing audio that is deliberately generated by text-to-speech, voice conversion, audio replay, etc.However, it has been shown that high-performance antispoofing models are vulnerable to adversarial attacks.Adversarial attacks, that are indistinguishable from original data but result in the incorrect predictions, are dangerous for antispoofing models and not in dispute we should detect them at any cost.To explore this issue, we proposed to employ Mockingjay, a self-supervised learning based model, to protect antispoofing models against adversarial attacks in the black-box scenario.Self-supervised learning models are effective in improving downstream task performance like phone classification or ASR.However, their effect in defense for adversarial attacks has not been explored yet.In this work, we explore the robustness of self-supervised learned high-level representations by using them in the defense against adversarial attacks.A layerwise noise to signal ratio (LNSR) is proposed to quantize and measure the effectiveness of deep models in countering adversarial noise.Experimental results on the ASVspoof 2019 dataset demonstrate that high-level representations extracted by Mockingjay can prevent the transferability of adversarial examples, and successfully counter black-box attacks.

Self-supervised Audio Transformers (SAT) enable great success in many downstream speech applications like ASR, but how they work has not been widely explored yet.In this work, we present multiple strategies for the analysis of attention mechanisms in SAT.We categorize attentions into explainable categories, where we discover each category possesses its own unique functionality.We provide a visualization tool for understanding multi-head self-attention, importance ranking strategies for identifying critical attention, and attention refinement techniques to improve model performance.

Score function-based natural language generation (NLG) approaches such as REINFORCE, in general, suffer from low sample efficiency and training instability problems. This is mainly due to the non-differentiable nature of the discrete space sampling and thus these methods have to treat the discriminator as a black box and ignore the gradient information. To improve the sample efficiency and reduce the variance of REINFORCE, we propose a novel approach, TaylorGAN, which augments the gradient estimation by off-policy update and the first-order Taylor expansion. This approach enables us to train NLG models from scratch with smaller batch size --- without maximum likelihood pre-training, and outperforms existing GAN-based methods on multiple metrics of quality and diversity.

High-performance spoofing countermeasure systems for automatic speaker verification (ASV) have been proposed in the ASVspoof 2019 challenge. However, the robustness of such systems under adversarial attacks has not been studied yet. In this paper, we investigate the vulnerability of spoofing countermeasures for ASV under both white-box and black-box adversarial attacks with the fast gradient sign method (FGSM) and the projected gradient descent (PGD) method. We implement high-performing countermeasure models in the ASVspoof 2019 challenge and conduct adversarial attacks on them. We compare performance of black-box attacks across spoofing countermeasure models with different network architectures and different amount of model parameters. The experimental results show that all implemented countermeasure models are vulnerable to FGSM and PGD attacks under the scenario of white-box attack. The more dangerous black-box attacks also prove to be effective by the experimental results.

Tsung-Yuan Hsu, Chi-Liang Liu, Hung-yi Lee. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Hong-Ren Mao, Hung-Yi Lee. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Yi-Lin Tuan, Yun-Nung Chen, Hung-yi Lee. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Yaushian Wang, Hung-Yi Lee, Yun-Nung Chen. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Spoken question answering (SQA) is challenging due to complex reasoning on top of the spoken documents. The recent studies have also shown the catastrophic impact of automatic speech recognition (ASR) errors on SQA. Therefore, this work proposes to mitigate the ASR errors by aligning the mismatch between ASR hypotheses and their corresponding reference transcriptions. An adversarial model is applied to this domain adaptation task, which forces the model to learn domain-invariant features the QA model can effectively utilize in order to improve the SQA results. The experiments successfully demonstrate the effectiveness of our proposed model, and the results are better than the previous best model by 2% EM score.

In this paper we proposed a novel Adversarial Training (AT) approach for end-to-end speech recognition using a Criticizing Language Model (CLM). In this way the CLM and the automatic speech recognition (ASR) model can challenge and learn from each other iteratively to improve the performance. Since the CLM only takes the text as input, huge quantities of unpaired text data can be utilized in this approach within end-to-end training. Moreover, AT can be applied to any end-to-end ASR model using any deep-learning-based language modeling frameworks, and compatible with any existing end-to-end decoding method. Initial results with an example experimental setup demonstrated the proposed approach is able to gain consistent improvements efficiently from auxiliary text data under different scenarios.

Speech-to-text translation (ST) refers to transforming the audio in source language to the text in target language. Mainstream solutions for such tasks are to cascade automatic speech recognition with machine translation, for which the transcriptions of the source language are needed in training. End-to-end approaches for ST tasks have been investigated because of not only technical interests such as to achieve globally optimized solution, but the need for ST tasks for the many source languages worldwide which do not have written form. In this paper, we propose a new end-to-end ST framework with two decoders to handle the relatively deeper relationships between the source language audio and target language text. The first-pass decoder generates some useful latent representations, and the second-pass decoder then integrates the output of both the encoder and the first-pass decoder to generate the text translation in target language. Only paired source language audio and target language text are used in training. Preliminary experiments on several language pairs showed improved performance, and offered some initial analysis.

Most state-of-the-art dialogue state tracking (DST) methods infer the dialogue state based on ground-truth transcriptions of utterances. In real-world situations, utterances are transcribed by automatic speech recognition (ASR) systems, which output the n-best candidate transcriptions (hypotheses). In certain noisy environments, the best transcription is often imperfect, severely influencing DST accuracy and possibly causing the dialogue system to stall or loop. The missed or misrecognized words can often be found in the runner-up candidate transcriptions from 2 to n, which could be used to improve accuracy of DST. However, looking beyond the top-ranked ASR results poses a dilemma: going too far may introduce noise, while not going far enough may not uncover any useful information. In this paper, we propose a novel approach to automatically determine the optimal time to stop reexamining runner-up ASR transcriptions based on deep reinforcement learning. Our method outperforms the baseline system, which uses only the top-1 ASR result, by 3.1%. Then, we select the dialogue rounds with the top-10 largest word error rate (WER), our method can improve DST accuracy by 15.4%, which is five times the overall improvement rate (3.1%). This improvement was expected because our proposed method is able to select informative ASR results at any rank.

Recent work has shown that exploiting relations between labels improves the performance of multi-label classification. We propose a novel framework based on generative adversarial networks (GANs) to model label dependency. The discriminator learns to model label dependency by discriminating real and generated label sets. To fool the discriminator, the classifier, or generator, learns to generate label sets with dependencies close to real data. Extensive experiments and comparisons on two large-scale image classification benchmark datasets (MS-COCO and NUS-WIDE) show that the discriminator improves generalization ability for different kinds of models.

Humans can imagine a scene from a sound. We want machines to do so by using conditional generative adversarial networks (GANs). By applying the techniques including spectral norm, projection discriminator and auxiliary classifier, compared with naive conditional GAN, the model can generate images with better quality in terms of both subjective and objective evaluations. Almost three-fourth of people agree that our model have the ability to generate images related to sounds. By inputting different volumes of the same sound, our model output different scales of changes based on the volumes, showing that our model truly knows the relationship between sounds and images to some extent.

Code-switching is about dealing with alternative languages in speech or text.It is partially speaker-dependent and domainrelated, so completely explaining the phenomenon by linguistic rules is challenging.Compared to most monolingual tasks, insufficient data is an issue for code-switching.To mitigate the issue without expensive human annotation, we proposed an unsupervised method for code-switching data augmentation.By utilizing a generative adversarial network, we can generate intra-sentential code-switching sentences from monolingual sentences.We applied the proposed method on two corpora, and the result shows that the generated code-switching sentences improve the performance of code-switching language models.

This paper focuses on using voice conversion (VC) to improve the speech intelligibility of surgical patients who have had parts of their articulators removed. Due to the difficulty of data collection, VC without parallel data is highly desired. Although techniques for unparallel VC, for example, CycleGAN, have been developed, they usually focus on transforming the speaker identity, and directly transforming the speech of one speaker to that of another speaker and as such do not address the task here. In this paper, we propose a new approach for unparallel VC. The proposed approach transforms impaired speech to normal speech while preserving the linguistic content and speaker characteristics. To our knowledge, this is the first end-to-end GAN-based unsupervised VC model applied to impaired speech. The experimental results show that the proposed approach outperforms CycleGAN.

End-to-end text-to-speech (TTS) has shown great success on large quantities of paired text plus speech data.However, laborious data collection remains difficult for at least 95% of the languages over the world, which hinders the development of TTS in different languages.In this paper, we aim to build TTS systems for such low-resource (target) languages where only very limited paired data are available.We show such TTS can be effectively constructed by transferring knowledge from a high-resource (source) language.Since the model trained on source language cannot be directly applied to target language due to input space mismatch, we propose a method to learn a mapping between source and target linguistic symbols.Benefiting from this learned mapping, pronunciation information can be preserved throughout the transferring procedure.Preliminary experiments show that we only need around 15 minutes of paired data to obtain a relatively good TTS system.Furthermore, analytic studies demonstrated that the automatically discovered mapping correlate well with the phonetic expertise.

Recently, voice conversion (VC) without parallel data has been successfully adapted to multi-target scenario in which a single model is trained to convert the input voice to many different speakers.However, such model suffers from the limitation that it can only convert the voice to the speakers in the training data, which narrows down the applicable scenario of VC.In this paper, we proposed a novel one-shot VC approach which is able to perform VC by only an example utterance from source and target speaker respectively, and the source and target speaker do not even need to be seen during training.This is achieved by disentangling speaker and content representations with instance normalization (IN).Objective and subjective evaluation shows that our model is able to generate the voice similar to target speaker.In addition to the performance measurement, we also demonstrate that this model is able to learn meaningful speaker representations without any supervision.

In this study, we propose a novel noise adaptive speech enhancement (SE) system, which employs a domain adversarial training (DAT) approach to tackle the issue of a noise type mismatch between the training and testing conditions. Such a mismatch is a critical problem in deep-learning-based SE systems. A large mismatch may cause a serious performance degradation to the SE performance. Because we generally use a well-trained SE system to handle various unseen noise types, a noise type mismatch commonly occurs in real-world scenarios. The proposed noise adaptive SE system contains an encoder-decoder-based enhancement model and a domain discriminator model. During adaptation, the DAT approach encourages the encoder to produce noise-invariant features based on the information from the discriminator model and consequentially increases the robustness of the enhancement model to unseen noise types. Herein, we regard stationary noises as the source domain (with the ground truth of clean speech) and non-stationary noises as the target domain (without the ground truth). We evaluated the proposed system on TIMIT sentences. The experiment results show that the proposed noise adaptive SE system successfully provides significant improvements in PESQ (19.0%), SSNR (39.3%), and STOI (27.0%) over the SE system without an adaptation.

We present an unsupervised end-to-end training scheme where we discover discrete subword units from speech without using any labels. The discrete subword units are learned under an ASR-TTS autoencoder reconstruction setting, where an ASR-Encoder is trained to discover a set of common linguistic units given a variety of speakers, and a TTS-Decoder trained to project the discovered units back to the designated speech. We propose a discrete encoding method, Multilabel-Binary Vectors (MBV), to make the ASR-TTS autoencoder differentiable. We found that the proposed encoding method offers automatic extraction of speech content from speaker style, and is sufficient to cover full linguistic content in a given language. Therefore, the TTS-Decoder can synthesize speech with the same content as the input of ASR-Encoder but with different speaker characteristics, which achieves voice conversion (VC). We further improve the quality of VC using adversarial training, where we train a TTS-Patcher that augments the output of TTS-Decoder. Objective and subjective evaluations show that the proposed approach offers strong VC results as it eliminates speaker identity while preserving content within speech. In the ZeroSpeech 2019 Challenge, we achieved outstanding performance in terms of low bitrate.

Speech separation has been very successful with deep learning techniques.Substantial effort has been reported based on approaches over spectrogram, which is well known as the standard time-and-frequency cross-domain representation for speech signals.It is highly correlated to the phonetic structure of speech, or "how the speech sounds" when perceived by human, but primarily frequency domain features carrying temporal behaviour.Very impressive work achieving speech separation over time domain was reported recently, probably because waveforms in time domain may describe the different realizations of speech in a more precise way than spectrogram.In this paper, we propose a framework properly integrating the above two directions, hoping to achieve both purposes.We construct a time-and-frequency feature map by concatenating the 1-dim convolution encoded feature map (for time domain) and the spectrogram (for frequency domain), which was then processed by an embedding network and clustering approaches very similar to those used in time and frequency domain prior works.In this way, the information in the time and frequency domains, as well as the interactions between them, can be jointly considered during embedding and clustering.Very encouraging results (state-of-the-art to our knowledge) were obtained with WSJ0-2mix dataset in preliminary experiments.

In text, word2vec transforms each word into a fixed-size vector used as the basic component in applications of natural language processing. Given a large collection of unannotated audio, audio word2vec can also be trained in an unsupervised way using a sequence-to-sequence autoencoder (SA). These vector representations are shown to effectively describe the sequential phonetic structures of the audio segments. In this paper, we further extend this research in the following two directions. First, we disentangle phonetic information and speaker information from the SA vector representations. Second, we extend audio word2vec from the word level to the utterance level by proposing a new segmental audio word2vec in which unsupervised spoken word boundary segmentation and audio word2vec are jointly learned and mutually enhanced, and utterances are directly represented as sequences of vectors carrying phonetic information. This is achieved by means of a segmental sequence-to-sequence autoencoder, in which a segmentation gate trained with reinforcement learning is inserted in the encoder.

A user can scan through a text easily, but it is not the case for spoken content, because they cannot be directly displayed on-screen. As a result, accessing large collections of spoken content is much more difficult and time-consuming than doing so for the text content. It would therefore be helpful to develop machines that understand spoken content. In this paper, we propose two new tasks for machine comprehension of spoken content. The first is a listening comprehension test for TOEFL, a challenging academic English examination for English learners who are not the native English speakers. We show that the proposed model outperforms the naive approaches and other neural network based models by exploiting the hierarchical structures of natural languages and the selective power of attention mechanism. For the second listening comprehension task - spoken SQuAD - we find that speech recognition errors severely impair machine comprehension; we propose the use of subword units to mitigate the impact of these errors.

Sequence generative adversarial networks (SeqGAN) have been used to improve conditional sequence generation tasks, for example, chit-chat dialogue generation. To stabilize the training of SeqGAN, Monte Carlo tree search (MCTS) or reward at every generation step (REGS) is used to evaluate the goodness of a generated subsequence. MCTS is computationally intensive, but the performance of REGS is worse than MCTS. In this paper, we propose stepwise GAN (StepGAN), in which the discriminator is modified to automatically assign scores quantifying the goodness of each subsequence at every generation step. StepGAN has significantly less computational costs than MCTS. We demonstrate that StepGAN outperforms previous GAN-based methods on both synthetic experiment and chit-chat dialogue generation.

Auto-encoders compress input data into a latent-space representation and reconstruct the original data from the representation.This latent representation is not easily interpreted by humans.In this paper, we propose training an auto-encoder that encodes input text into human-readable sentences, and unpaired abstractive summarization is thereby achieved.The auto-encoder is composed of a generator and a reconstructor.The generator encodes the input text into a shorter word sequence, and the reconstructor recovers the generator input from the generator output.To make the generator output human-readable, a discriminator restricts the output of the generator to resemble human-written sentences.By taking the generator output as the summary of the input text, abstractive summarization is achieved without document-summary pairs as training data.Promising results are shown on both English and Chinese corpora.

Retrieving spoken content with spoken queries, or query-by-example spoken term detection (STD), is attractive because it makes possible the matching of signals directly on the acoustic level without transcribing them into text. Here, we propose an end-to-end query-by-example STD model based on an attention-based multi-hop network, whose input is a spoken query and an audio segment containing several utterances; the output states whether the audio segment includes the query. The model can be trained in either a supervised scenario using labeled data, or in an unsupervised fashion. In the supervised scenario, we find that the attention mechanism and multiple hops improve performance, and that the attention weights indicate the time span of the detected terms. In the unsupervised setting, the model mimics the behavior of DTW, and it performs as well as DTW but with a lower run-time complexity.

Conventional seq2seq chatbot models only try to find the sentences with the highest probabilities conditioned on the input sequences, without considering the sentiment of the output sentences. Some research works trying to modify the sentiment of the output sequences were reported. In this paper, we propose five models to scale or adjust the sentiment of the chatbot response: persona-based model, reinforcement learning, plug and play model, sentiment transformation network and cycleGAN, all based on the conventional seq2seq model. We also develop two evaluation metrics to estimate if the responses are reasonable given the input. These metrics together with other two popularly used metrics were used to analyze the performance of the five proposed models on different aspects, and reinforcement learning and cycleGAN were shown to be very attractive. The evaluation metrics were also found to be well correlated with human evaluation.

This paper proposes a domain independent approach for extracting key terms from spoken content based on context and term location information, or the sentence structures. Once it is trained with data of enough different domains, it is able to extract key terms in other unseen domains. This is obviously very attractive because of the unlimited number of domains over the Internet. Its performance degrades only very slightly with recognition errors, so very useful for spoken content. The basic idea here is that the sentence structures or context and term location information are in general domain independent, and remain essentially unchanged with recognition errors. For example, the fact that the key term for the sentence “The subject of this article is primarily about neural networks” is “neural networks” can be extended to any other unseen term other than “neural networks” in any other unseen domain, and this is more or less preserved under recognition errors. In the experiments a model trained with data for five different domains can extract key terms from data in the sixth unseen domain with very good performance.

Audio Word2Vec offers vector representations of fixed dimensionality for variable-length audio segments using Sequence to-sequence Autoencoder (SA). These vector representations are shown to describe the sequential phonetic structures of the audio segments to a good degree, with real world applications such as spoken term detection (STD). This paper examines the capability of language transfer of Audio Word2Vec. We train SA from one language (source language) and use it to extract the vector representation of the audio segments of another language (target language). We found that SA can still catch the phonetic structure from the audio segments of the target language if the source and target languages are similar. In STD, we obtain the vector representations from the SA learned from a large amount of source language data, and found them surpass the representations from naive encoder and SA directly learned from a small amount of target language data. The result shows that it is possible to learn Audio Word2Vec model from high-resource languages and use it on low-resource languages. This further expands the usability of Audio Word2Vec.

While Word2Vec represents words (in text) as vectors carrying semantic information, audio Word2Vec was shown to be able to represent signal segments of spoken words as vectors carrying phonetic structure information. Audio Word2Vec can be trained in an unsupervised way from an unlabeled corpus, except the word boundaries are needed. In this paper, we extend audio Word2Vec from word-level to utterance-level by proposing a new segmental audio Word2Vec, in which unsupervised spoken word boundary segmentation and audio Word2Vec are jointly learned and mutually enhanced, so an utterance can be directly represented as a sequence of vectors carrying phonetic structure information. This is achieved by a segmental sequence-to-sequence autoencoder (SSAE), in which a segmentation gate trained with reinforcement learning is inserted in the encoder. Experiments on English, Czech, French and German show very good performance in both unsupervised spoken word segmentation and spoken term detection applications (significantly better than frame-based DTW).

Recently, cycle-consistent adversarial network (Cycle-GAN) has been successfully applied to voice conversion to a different speaker without parallel data, although in those approaches an individual model is needed for each target speaker.In this paper, we propose an adversarial learning framework for voice conversion, with which a single model can be trained to convert the voice to many different speakers, all without parallel data, by separating the speaker characteristics from the linguistic content in speech signals.An autoencoder is first trained to extract speaker-independent latent representations and speaker embedding separately using another auxiliary speaker classifier to regularize the latent representation.The decoder then takes the speaker-independent latent representation and the target speaker embedding as the input to generate the voice of the target speaker with the linguistic content of the source utterance.The quality of decoder output is further improved by patching with the residual signal produced by another pair of generator and discriminator.A target speaker set size of 20 was tested in the preliminary experiments, and very good voice quality was obtained.Conventional voice conversion metrics are reported.We also show that the speaker information has been properly reduced from the latent representations.

User-machine interaction is crucial for information retrieval, especially for spoken content retrieval, because spoken content is difficult to browse, and speech recognition has a high degree of uncertainty. In interactive retrieval, the machine takes different actions to interact with the user to obtain better retrieval results; here it is critical to select the most efficient action. In previous work, deep Q-learning techniques were proposed to train an interactive retrieval system but rely on a hand-crafted user simulator; building a reliable user simulator is difficult. In this paper, we further improve the interactive spoken content retrieval framework by proposing a learnable user simulator which is jointly trained with interactive retrieval system, making the hand-crafted user simulator unnecessary. The experimental results show that the learned simulated users not only achieve larger rewards than the hand-crafted ones but act more like real users.

Reading comprehension has been widely studied. One of the most representative reading comprehension tasks is Stanford Question Answering Dataset (SQuAD), on which machine is already comparable with human. On the other hand, accessing large collections of multimedia or spoken content is much more difficult and time-consuming than plain text content for humans. It's therefore highly attractive to develop machines which can automatically understand spoken content. In this paper, we propose a new listening comprehension task - Spoken SQuAD. On the new task, we found that speech recognition errors have catastrophic impact on machine comprehension, and several approaches are proposed to mitigate the impact.

Unsupervised discovery of acoustic tokens from audio corpora without annotation and learning vector representations for these tokens have been widely studied.Although these techniques have been shown successful in some applications such as query-byexample Spoken Term Detection (STD), the lack of mapping relationships between these discovered tokens and real phonemes have limited the down-stream applications.This paper represents probably the first attempt towards the goal of completely unsupervised phoneme recognition, or mapping audio signals to phoneme sequences without phoneme-labeled audio data.The basic idea is to cluster the embedded acoustic tokens and learn the mapping between the cluster sequences and the unknown phoneme sequences with a Generative Adversarial Network (GAN).An unsupervised phoneme recognition accuracy of 36% was achieved in the preliminary experiments.

Yu-An Chung, Hung-Yi Lee, James Glass. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Word embedding or Word2Vec has been successful in offering semantics for text words learned from the context of words. Audio Word2Vec was shown to offer phonetic structures for spoken words (signal segments for words) learned from signals within spoken words. This paper proposes a two-stage framework to perform phonetic-and-semantic embedding on spoken words considering the context of the spoken words. Stage 1 performs phonetic embedding with speaker characteristics disentangled. Stage 2 then performs semantic embedding in addition. We further propose to evaluate the phonetic-and-semantic nature of the audio embeddings obtained in Stage 2 by parallelizing with text embeddings. In general, phonetic structure and semantics inevitably disturb each other. For example the words “brother” and “sister” are close in semantics but very different in phonetic structure, while the words “brother” and “bother” are in the other way around. But phonetic-and-semantic embedding is attractive, as shown in the initial experiments on spoken document retrieval. Not only spoken documents including the spoken query can be retrieved based on the phonetic structures, but spoken documents semantically related to the query but not including the query can also be retrieved based on the semantics.

Reading comprehension by machine has been widely studied, but machine comprehension of spoken content is still a less investigated problem. In this paper, we release Open-Domain Spoken Question Answering Dataset (ODSQA) with more than three thousand questions. To the best of our knowledge, this is the largest real SQA dataset. On this dataset, we found that ASR errors have catastrophic impact on SQA. To mitigate the effect of ASR errors, subword units are involved, which brings consistent improvements over all the models. We further found that data augmentation on text-based QA training examples can improve SQA.

End-to-end TTS model can directly take an utterance as reference, and generate speech from the text with prosody and speaker characteristics similar to the reference utterance. Ideally, the transcription of reference utterance does not need to match the text to be synthesized, so unsupervised style transfer can be achieved. However, in the previous model, because only the matched text and speech are used in training, given unmatched text and speech during testing would make the model synthesize blurry speech. In this paper, we propose to mitigate the problem by using the unmatched text and speech during training, and using the ASR accuracy of an end-to-end ASR model to guide the training procedure. The experimental results show that with the guidance of end-to-end ASR, both the ASR accuracy (objective evaluation) and the listener preference (subjective evaluation) of the speech generated by TTS model are improved. Moreover, we propose attention consistency loss as regularization, which is shown to accelerate the training.

Speaking rate refers to the average number of phonemes within some unit time, while the rhythmic patterns refer to duration distributions for realizations of different phonemes within different phonetic structures. Both are key components of prosody in speech, which is different for different speakers. Models like cycle-consistent adversarial network (Cycle-GAN) and variational auto-encoder (VAE) have been successfully applied to voice conversion tasks without parallel data. However, due to the neural network architectures and feature vectors chosen for these approaches, the length of the predicted utterance has to be fixed to that of the input utterance, which limits the flexibility in mimicking the speaking rates and rhythmic patterns for the target speaker. On the other hand, sequence-to-sequence learning model was used to remove the above length constraint, but parallel training data are needed. In this paper, we propose an approach utilizing sequence-to-sequence model trained with unsupervised Cycle-GAN to perform the transformation between the phoneme posteriorgram sequences for different speakers. In this way, the length constraint mentioned above is removed to offer rhythm-flexible voice conversion without requiring parallel data. Preliminary evaluation on two datasets showed very encouraging results.

For text content retrieval, the user can easily scan through and select from a list of retrieved items. This is impossible for spoken content retrieval, because the retrieved items are not easily displayed on-screen. In addition, due to the high degree of uncertainty for speech recognition, retrieval results can be very noisy. One way to counter such difficulties is through user-machine interaction. The machine can take different actions to interact with the user to obtain better retrieval results before showing them to the user. For example, the machine can request extra information from the user, return a list of topics for the user to select from, and so on. In this paper, we propose using deep-Q-network (DQN) to determine the machine actions for interactive spoken content retrieval. DQN bypasses the need to estimate hand-crafted states, and directly determines the best action based on the present retrieval results even without any human knowledge. It is shown to achieve significantly better performance as compared with the previous hand-crafted states. We further find that double DQN and dueling DQN improve the naive version.

We apply sequence-to-sequence model to mitigate the impact of speech recognition errors on open domain end-to-end dialog generation. We cast the task as a domain adaptation problem where ASR transcriptions and original texts are in two different domains. In this paper, our proposed model includes two individual encoders for each domain data and make their hidden states similar to ensure the decoder predict the same dialog text. The method demonstrates that the sequence-to-sequence model can learn the ASR transcriptions and original text pair having the same meaning and eliminate the speech recognition errors. Experimental results on Cornell movie dialog dataset demonstrate that the domain adaption system help the spoken dialog system generate more similar responses with the original text answers.

Video captioning has been widely researched. Most related work takes into account only visual content in generating descriptions. However, auditory content such as human speech or environmental sounds contains rich information for describing scenes, but has yet to be widely explored for video captions. Here, we experiment with different ways to use this auditory content in videos, and demonstrate improved caption generation in terms of popular evaluation methods such as BLEU, CIDEr, and METEOR. We also measure the semantic similarities between generated captions and human-provided ground truth using sentence embeddings, and find that good use of multi-modal contents helps the machine to generate captions that are more semantically related to the ground truth. When analyzing the generated sentences, we find some ambiguous situations for which visual-only models yield incorrect results but that are resolved by approaches that take into account auditory cues.

Distributed word representations have been shown to be very useful in various natural language processing (NLP) application tasks. These word vectors learned from huge corpora very often carry both semantic and syntactic information of words. However, it is well known that each individual user has his own language patterns because of different factors such as interested topics, friend groups, social activities, wording habits, etc., which may imply some kind of personalized semantics. With such personalized semantics, the same word may imply slightly differently for different users. For example, the word “Cappuccino” may imply “Leisure”, “Joy”, “Excellent” for a user enjoying coffee, by only a kind of drink for someone else. Such personalized semantics of course cannot be carried by the standard universal word vectors trained with huge corpora produced by many people. In this paper, we propose a framework to train different personalized word vectors for different users based on the very successful continuous skip-gram model using the social network data posted by many individual users. In this framework, universal background word vectors are first learned from the background corpora, and then adapted by the personalized corpus for each individual user to learn the personalized word vectors. We use two application tasks to evaluate the quality of the personalized word vectors obtained in this way, the user prediction task and the sentence completion task. These personalized word vectors were shown to carry some personalized semantics and offer improved performance on these two evaluation tasks.

In this paper, we propose new methods to learn Chinese word representations.Chinese characters are composed of graphical components, which carry rich semantics.It is common for a Chinese learner to comprehend the meaning of a word from these graphical components.As a result, we propose models that enhance word representations by character glyphs.The character glyph features are directly learned from the bitmaps of characters by convolutional auto-encoder(convAE), and the glyph features improve Chinese word representations which are already enhanced by character embeddings.Another contribution in this paper is that we created several evaluation datasets in traditional Chinese and made them public.

It is crucial for language models to model long-term dependency in word sequences, which can be achieved to some good extent by recurrent neural network (RNN) based language models with long short-term memory (LSTM) units. To accurately model the sophisticated long-term information in human languages, large memory in language models is necessary. However, the size of RNN-based language models cannot be arbitrarily increased because the computational resources required and the model complexity will also be increase accordingly, due to the limitation of the structure. To overcome this problem, inspired from Neural Turing Machine and Memory Network, we equip RNN-based language models with controllable external memory. With a learnable memory controller, the size of the external memory is independent to the number of model parameters, so the proposed language model can have larger memory without increasing the parameters. In the experiments, the proposed model yielded lower perplexities than RNN-based language models with LSTM units on both English and Chinese corpora.

It is well known that recognizers personalized to each user are much more effective than user-independent recognizers. With the popularity of smartphones today, although it is not difficult to collect a large set of audio data for each user, it is difficult to transcribe it. However, it is now possible to automatically discover acoustic tokens from unlabeled personal data in an unsupervised way. We therefore propose a multi-task deep learning framework called a phoneme-token deep neural network (PTDNN), jointly trained from unsupervised acoustic tokens discovered from unlabeled data and very limited transcribed data for personalized acoustic modeling. We term this scenario “weakly supervised”. The underlying intuition is that the high degree of similarity between the HMM states of acoustic token models and phoneme models may help them learn from each other in this multi-task learning framework. Initial experiments performed over a personalized audio data set recorded from Facebook posts demonstrated that very good improvements can be achieved in both frame accuracy and word accuracy over popularly-considered baselines such as fDLR, speaker code and lightly supervised adaptation. This approach complements existing speaker adaptation approaches and can be used jointly with such techniques to yield improved results.

Connectionist temporal classification (CTC) is a powerful approach for sequence-to-sequence learning, and has been popularly used in speech recognition.The central ideas of CTC include adding a label "blank" during training.With this mechanism, CTC eliminates the need of segment alignment, and hence has been applied to various sequence-to-sequence learning problems.In this work, we applied CTC to abstractive summarization for spoken content.The "blank" in this case implies the corresponding input data are less important or noisy; thus it can be ignored.This approach was shown to outperform the existing methods in term of ROUGE scores over Chinese Gigaword and MATBN corpora.This approach also has the nice property that the ordering of words or characters in the input documents can be better preserved in the generated summaries.

In this paper we analyze the gate activation signals inside the gated recurrent neural networks, and find the temporal structure of such signals is highly correlated with the phoneme boundaries. This correlation is further verified by a set of experiments for phoneme segmentation, in which better results compared to standard approaches were obtained.

With the popularity of mobile devices, personalized speech recognizers have become more attainable and are highly attractive. Since each mobile device is used primarily by a single user, it is possible to have a personalized recognizer that well matches the characteristics of the individual user. Although acoustic model personalization has been investigated for decades, much less work has been reported on personalizing language models, presumably because of the difficulties in collecting sufficient personalized corpora. In this paper, we propose a general framework for personalizing recurrent-neural-network-based language models (RNNLMs) using data collected from social networks, including the posts of many individual users and friend relationships among the users. Two major directions for this are model-based and feature-based RNNLM personalization. In model-based RNNLM personalization, the RNNLM parameters are fine-tuned to an individual user's wording patterns by incorporating social texts posted by the target user and his or her friends. For the feature-based approach, the RNNLM model parameters are fixed across users, but the RNNLM input features are instead augmented with personalized information. Both approaches not only drastically reduce the model perplexity, but also moderately reduce word error rates in n-best rescoring tests.

The vector representations of fixed dimensionality for words (in text) offered by Word2Vec have been shown to be very useful in many application scenarios, in particular due to the semantic information they carry.This paper proposes a parallel version, the Audio Word2Vec.It offers the vector representations of fixed dimensionality for variable-length audio segments.These vector representations are shown to describe the sequential phonetic structures of the audio segments to a good degree, with very attractive real world applications such as query-by-example Spoken Term Detection (STD).In this STD application, the proposed approach significantly outperformed the conventional Dynamic Time Warping (DTW) based approaches at significantly lower computation requirements.We propose unsupervised learning of Audio Word2Vec from audio data without human annotation using Sequence-to-sequence Autoencoder (SA).SA consists of two RNNs equipped with Long Short-Term Memory (LSTM) units: the first RNN (encoder) maps the input audio sequence into a vector representation of fixed dimensionality, and the second RNN (decoder) maps the representation back to the input audio sequence.The two RNNs are jointly trained by minimizing the reconstruction error.Denoising Sequence-to-sequence Autoencoder (DSA) is further proposed offering more robust learning.

Recurrent neural network architectures combining with attention mechanism, or neural attention model, have shown promising performance recently for the tasks including speech recognition, image caption generation, visual question answering and machine translation.In this paper, neural attention model is applied on two sequence labeling tasks, dialogue act detection and key term extraction.In the sequence labeling tasks, the model input is a sequence, and the output is the label of the input sequence.The major difficulty of sequence labeling is that when the input sequence is long, it can include many noisy or irrelevant part.If the information in the whole sequence is treated equally, the noisy or irrelevant part may degrade the classification performance.The attention mechanism is helpful for sequence classification task because it is capable of highlighting important part among the entire sequence for the classification task.The experimental results show that with the attention mechanism, discernible improvements were achieved in the sequence labeling task considered here.The roles of the attention mechanism in the tasks are further analyzed and visualized in this paper.

Multimedia or spoken content presents more attractive information than plain text content, but it's more difficult to display on a screen and be selected by a user. As a result, accessing large collections of the former is much more difficult and time-consuming than the latter for humans. It's highly attractive to develop a machine which can automatically understand spoken content and summarize the key information for humans to browse over. In this endeavor, we propose a new task of machine comprehension of spoken content. We define the initial goal as the listening comprehension test of TOEFL, a challenging academic English examination for English learners whose native language is not English. We further propose an Attention-based Multi-hop Recurrent Neural Network (AMRNN) architecture for this task, achieving encouraging results in the initial tests. Initial results also have shown that word-level attention is probably more robust than sentence-level attention for this task with ASR errors.

User-machine interaction is important for spoken content retrieval.For text content retrieval, the user can easily scan through and select on a list of retrieved item.This is impossible for spoken content retrieval, because the retrieved items are difficult to show on screen.Besides, due to the high degree of uncertainty for speech recognition, the retrieval results can be very noisy.One way to counter such difficulties is through user-machine interaction.The machine can take different actions to interact with the user to obtain better retrieval results before showing to the user.The suitable actions depend on the retrieval status, for example requesting for extra information from the user, returning a list of topics for user to select, etc.In our previous work, some hand-crafted states estimated from the present retrieval results are used to determine the proper actions.In this paper, we propose to use Deep-Q-Learning techniques instead to determine the machine actions for interactive spoken content retrieval.Deep-Q-Learning bypasses the need for estimation of the hand-crafted states, and directly determine the best action base on the present retrieval status even without any human knowledge.It is shown to achieve significantly better performance compared with the previous hand-crafted states.

Multimedia or spoken content presents more attractive information than plain text content, but the former is more difficult to display on a screen and be selected by a user. As a result, accessing large collections of the former is much more difficult and time-consuming than the latter for humans. It's therefore highly attractive to develop machines which can automatically understand spoken content and summarize the key information for humans to browse over. In this endeavor, a new task of machine comprehension of spoken content was proposed recently. The initial goal was defined as the listening comprehension test of TOEFL, a challenging academic English examination for English learners whose native languages are not English. An Attention-based Multi-hop Recurrent Neural Network (AMRNN) architecture was also proposed for this task, which considered only the sequential relationship within the speech utterances. In this paper, we propose a new Hierarchical Attention Model (HAM), which constructs multi-hopped attention mechanism over tree-structured rather than sequential representations for the utterances. Improved comprehension performance robust with respect to ASR errors were obtained.

Headline generation for spoken content is important since spoken content is difficult to be shown on the screen and browsed by the user. It is a special type of abstractive summarization, for which the summaries are generated word by word from scratch without using any part of the original content. Many deep learning approaches for headline generation from text document have been proposed recently, all requiring huge quantities of training data, which is difficult for spoken document summarization. In this paper, we propose an ASR error modeling approach to learn the underlying structure of ASR error patterns and incorporate this model in an Attentive Recurrent Neural Network (ARNN) architecture. In this way, the model for abstractive headline generation for spoken content can be learned from abundant text data and the ASR data for some recognizers. Experiments showed very encouraging results and verified that the proposed ASR error model works well even when the input spoken content is recognized by a recognizer very different from the one the model learned from.

In this work we aim to discover high quality speech features and Linguistic units directly from unlabeled speech data in a zero resource scenario. The results are evaluated using the metrics and corpora proposed in the Zero Resource Speech Challenge organized at Interspeech 2015. A Multi-layered Acoustic Tokenizer (MAT) was proposed for automatic discovery of multiple sets of acoustic tokens from the given corpus. Each acoustic token set is specified by a set of hyperparameters that describe the model configuration. These sets of acoustic tokens carry different characteristics fof the given corpus and the language behind, thus can be mutually reinforced. The multiple sets of token labels are then used as the targets of a Multi-target Deep Neural Network (MDNN) trained on low-level acoustic features. Bottleneck features extracted from the MDNN are then used as the feedback input to the MAT and the MDNN itself in the next iteration. We call this iterative deep learning framework the Multi-layered Acoustic Tokenizing Deep Neural Network (MAT-DNN), which generates both high quality speech features for the Track 1 of the Challenge and acoustic tokens for the Track 2 of the Challenge. In addition, we performed extra experiments on the same corpora on the application of query-by-example spoken term detection. The experimental results showed the iterative deep learning framework of MAT-DNN improved the detection performance due to better underlying speech features and acoustic tokens.

In this paper we propose the Structured Deep Neural Network (structured DNN) as a structured and deep learning framework. This approach can learn to find the best structured object (such as a label sequence) given a structured input (such as a vector sequence) by globally considering the mapping relationships between the structures rather than item by item. When automatic speech recognition is viewed as a special case of such a structured learning problem, where we have the acoustic vector sequence as the input and the phoneme label sequence as the output, it becomes possible to comprehensively learn utterance by utterance as a whole, rather than frame by frame. Structured Support Vector Machine (structured SVM) was proposed to perform ASR with structured learning previously, but limited by the linear nature of SVM. Here we propose structured DNN to use nonlinear transformations in multi-layers as a structured and deep learning approach. This approach was shown to beat structured SVM in preliminary experiments on TIMIT.

With the popularity of mobile devices, personalized speech recognizer becomes more realizable today and highly attractive. Each mobile device is primarily used by a single user, so it's possible to have a personalized recognizer well matching to the characteristics of individual user. Although acoustic model personalization has been investigated for decades, much less work have been reported on personalizing language model, probably because of the difficulties in collecting enough personalized corpora. Previous work used the corpora collected from social networks to solve the problem, but constructing a personalized model for each user is troublesome. In this paper, we propose a universal recurrent neural network language model with user characteristic features, so all users share the same model, except each with different user characteristic features. These user characteristic features can be obtained by crowdsouring over social networks, which include huge quantity of texts posted by users with known friend relationships, who may share some subject topics and wording patterns. The preliminary experiments on Facebook corpus showed that this proposed approach not only drastically reduced the model perplexity, but offered very good improvement in recognition accuracy in n-best rescoring tests. This approach also mitigated the data sparseness problem for personalized language models.

The increasing popularity of Massive Open Online Courses (MOOCs) has resulted in huge number of courses available over the Internet. Typically, a learner can type a search query into the look-up window of a MOOC platform and receive a set of course suggestions. But it is difficult for the learner to select lectures out of those suggested courses and learn the desired information efficiently. In this paper, we propose to structure the lectures of the various suggested courses into a map (graph) for each query entered by the learner, indicating the lectures with very similar content and reasonable sequence order of learning. In this way the learner can define his own learning path on the map based on his interests and backgrounds, and learn the desired information from lectures in different courses without too much difficulties in minimum time. We propose a series of approaches for linking lectures of very similar content and predicting the prerequisites for this purpose. Preliminary results show that the proposed approaches have the potential to achieve the above goal.

Spoken content retrieval refers to directly indexing and retrieving spoken content based on the audio rather than text descriptions. This potentially eliminates the requirement of producing text descriptions for multimedia content for indexing and retrieval purposes, and is able to precisely locate the exact time the desired information appears in the multimedia. Spoken content retrieval has been very successfully achieved with the basic approach of cascading automatic speech recognition (ASR) with text information retrieval: after the spoken content is transcribed into text or lattice format, a text retrieval engine searches over the ASR output to find desired information. This framework works well when the ASR accuracy is relatively high, but becomes less adequate when more challenging real-world scenarios are considered, since retrieval performance depends heavily on ASR accuracy. This challenge leads to the emergence of another approach to spoken content retrieval: to go beyond the basic framework of cascading ASR with text retrieval in order to have retrieval performances that are less dependent on ASR accuracy. This overview article is intended to provide a thorough overview of the concepts, principles, approaches, and achievements of major technical contributions along this line of investigation. This includes five major directions: 1) Modified ASR for Retrieval Purposes: cascading ASR with text retrieval, but the ASR is modified or optimized for spoken content retrieval purposes; 2) Exploiting the Information not present in ASR outputs: to try to utilize the information in speech signals inevitably lost when transcribed into phonemes and words; 3) Directly Matching at the Acoustic Level without ASR: for spoken queries, the signals can be directly matched at the acoustic level, rather than at the phoneme or word levels, bypassing all ASR issues; 4) Semantic Retrieval of Spoken Content: trying to retrieve spoken content that is semantically related to the query, but not necessarily including the query terms themselves; 5) Interactive Retrieval and Efficient Presentation of the Retrieved Objects: with efficient presentation of the retrieved objects, an interactive retrieval process incorporating user actions may produce better retrieval results and user experiences.

Acoustic feature similarity between search results has been shown to be very helpful for the task of spoken term detection (STD). A graph-based re-ranking approach for STD has been proposed based on the concept that search results, which are acoustically similar to other results with higher confidence scores, should have higher scores themselves. In this approach, the similarity between all search results of a given term are considered as a graph, and the confidence scores of the search results propagate through this graph. Since this approach can improve STD results without any additional labelled data, it is especially suitable for STD on languages with limited amounts of annotated data. However, its performance has not been widely studied on benchmark corpora. In this paper, we investigate the effectiveness of the graph-based reranking approach on limited language data from the IARPA Babel program. Experiments on the low-resource languages, Assamese, Bengali and Lao, show that graph-based re-ranking improves STD systems using fuzzy matching, and lattices based on different kinds of units including words, subwords, and hybrids. Index Terms: Random Walk, Spoken Term Detection

In a text context, document/query expansion has proven very useful in retrieving objects semantically related to the query. However, when applying text-based techniques on spoken content, the inevitable recognition errors seriously degrade performance even when the retrieval process is performed over lattices. We propose the estimation of more accurate term distributions (or unigram language models) for the spoken documents by acoustic similarity graphs. In this approach, a graph is constructed for each term describing the acoustic similarity among all signal regions hypothesized to be the considered term. Score propagation based on a random walk over the graph offers more reliable scores of the term hypotheses, which in turn yield more accurate term distributions (or unigram language models). This approach was applied with the language modeling retrieval approach, including using document expansion based on latent topic analysis and query expansion with a query-regularized mixture model. We extend these approaches from words to subword n-grams, and the query expansion from document-level to utterance-level and from term-based to topic-based. Experiments performed on Mandarin broadcast news showed improved performance under almost all tested conditions.

It takes very long time to go through a complete online course. Without proper background, it is also difficult to understand retrieved spoken paragraphs. This paper therefore presents a new approach of spoken knowledge organization for course lectures for efficient personalized learning. Automatically extracted key terms are taken as the fundamental elements of the semantics of the course. Key term graph constructed by connecting related key terms forms the backbone of the global semantic structure. Audio/video signals are divided into multi-layer temporal structure including paragraphs, sections and chapters, each of which includes a summary as the local semantic structure. The interconnection between semantic structure and temporal structure together with spoken term detection jointly offer to the learners efficient ways to navigate across the course knowledge with personalized learning paths considering their personal interests, available time and background knowledge. A preliminary prototype system has also been successfully developed.

This paper presents an initial effort to retrieve semantically related spoken content in a completely unsupervised way. Unsupervised approaches of spoken content retrieval is attractive because the need for annotated data reasonably matched to the spoken content for training acoustic and language models can be bypassed. However, almost all such unsupervised approaches focus on spoken term detection, or returning the spoken segments containing the query, using either template matching techniques such as dynamic time warping (DTW) or model-based approaches. However, users usually prefer to retrieve all objects semantically related to the query, but not necessarily including the query terms. This paper proposes a different approach. We transcribe the spoken segments in the archive to be retrieved through into sequences of acoustic patterns automatically discovered in an unsupervised method. For an input query in spoken form, the top-N spoken segments from the archive obtained with the first-pass retrieval with DTW are taken as pseudo-relevant. The acoustic patterns frequently occurring in these segments are therefore considered as query-related and used for query expansion. Preliminary experiments performed on Mandarin broadcast news offered very encouraging results.

Supervised approaches can learn a spoken document summarizer generating high-quality summaries using a set of training examples matched to the domain of target documents. However, preparing a sufficient number of in-domain training examples is expensive. In this paper we propose an approach for unsupervised domain adaptation for spoken document summarization, so no in-domain training examples are needed. A summarizer is first learned from a set of out-of-domain training examples by a supervised summarization approach based on structured support vector machine, and this summarizer is used to generate a set of initial summaries for the target spoken documents. The target documents and their initial machine-generated summaries then serve as extra training examples for learning a new summarizer, which further updates the summaries of the target spoken documents. This process is continued iteratively to incrementally improve the summarizer for the target spoken documents. Moreover, extra approaches transforming the feature representations based on the data distribution in the target domain and augmenting the representations with an extra set of domain-specific features are also proposed. Encouraging results were obtained in summarizing Mandarin-English code-switching course lectures using training examples from Mandarin broadcast news.

Query expansion techniques were originally developed for text information retrieval in order to retrieve the documents not containing the query terms but semantically related to the query. This is achieved by assuming the terms frequently occurring in the top-ranked documents in the first-pass retrieval results to be query-related and using them to expand the query to do the second-pass retrieval. However, when this approach was used for spoken content retrieval, the inevitable recognition errors and the OOV problems in ASR make it difficult for many query-related terms to be included in the expanded query, and much of the information carried by the speech signal is lost during recognition and not recoverable. In this paper, we propose to use a second ASR engine based on acoustic patterns automatically discovered from the spoken archive used for retrieval. These acoustic patterns are discovered directly based on the signal characteristics, and therefore can compensate for the information lost during recognition to a good extent. When a text query is entered, the system generates the first-pass retrieval results based on the transcriptions of the spoken segments obtained via the conventional ASR. The acoustic patterns frequently occurring in the spoken segments ranked on top of the first-pass results are considered as query-related, and the spoken segments containing these query-related acoustic patterns are retrieved. In this way, even though some query-related terms are OOV or incorrectly recognized, the segments including these terms can still be retrieved by acoustic patterns corresponding to these terms. Preliminary experiments performed on Mandarin broadcast news offered very encouraging results.

Interactive retrieval is important for spoken content because the retrieved spoken items are not only difficult to be shown on the screen but also scanned and selected by the user, in addition to the speech recognition uncertainty. The user cannot playback and go through all the retrieved items to find out what he is looking for. Markov Decision Process (MDP) was used in a previous work to help the system take different actions to interact with the user based on an estimated retrieval performance, but the MDP state was represented by the less precise quantized retrieval performance metric. In this paper, we consider the retrieval performance metric as a continuous state variable in MDP and optimize the MDP by fitted value iteration (FVI).We also use query expansion with the language modeling retrieval framework to produce the next set of retrieval results. Improved performance was found in the preliminary experiments.

Speech recognition has become an important feature in smartphones in recent years. Different from traditional au-tomatic speech recognition, the speech recognition on smart-phones can take advantage of personalized language models to model the linguistic patterns and wording habits of a particu-lar smartphone owner better. Owing to the popularity of social networks in recent years, personal texts and messages are no longer inaccessible. However, data sparseness is still an un-solved problem. In this paper, we propose a three-step adapta-tion approach to personalize recurrent neural network language models (RNNLMs). We believe that its capability to model word histories as distributed representations of arbitrary length can help mitigate the data sparseness problem. Furthermore, we also propose additional user-oriented features to empower the RNNLMs with stronger capabilities for personalization. The experiments on a Facebook dataset showed that the proposed method not only drastically reduced the model perplexity in preliminary experiments, but also moderately reduced the word error rate in n-best rescoring tests.

Spoken term detection (STD) is a key technology for retrieval of spoken content, which will be very important to retrieve and browse multimedia content over the Internet. The discriminative capability of machine learning methods has recently been used to facilitate STD. This paper presents a new approach to improve STD using support vector machines (SVM) based on acoustic information. The concept of pseudo-relevance feedback (PRF) well used in the retrieval of text, image and video is used here. The basic idea of using PRF here is to assume some spoken segments in the first-pass retrieved results are relevant (or pseudo-relevant) and some others irrelevant (or pseudo-irrelevant), and take these segments as positive and negative examples to train a query-specific SVM. This SVM is then used for re-ranking the first-pass retrieved results, and only the re-ranked results are shown to the user. In this paper, feature vectors representing the spoken segments based on acoustic information to be used in SVM are considered and analyzed. Furthermore, conventionally in PRF the items with the highest and lowest scores in the first-pass retrieved results are respectively taken as pseudo-relevant and -irrelevant, but in this way some incorrect examples are inevitably included in the training data especially when the recognition accuracy is poor. Here we further propose an enhanced SVM which not only better selects positive/negative examples considering the reliability of the spoken segments, but emphasizes more on more reliable training examples by modifying the SVM formulation. Experiments on two different sets of spoken archives with different speaking styles and different levels of recognition accuracies demonstrated significant improvements offered by the proposed approaches.

This paper proposes an unsupervised two-stage approach to automatically extract keywords from spoken documents. In the first stage, for each candidate term we compute a topic coherence and term significance measure (TCS) based on probabilistic latent semantic analysis (PLSA) models. In the second stage, we take the candidate terms with highest and lowest TCS scores as positive and negative examples to train an SVM classifier in an unsupervised way using prosodic, lexical, and semantic features, and then classify the candidate keyword using this SVM classifier. The experiments with course lectures showed that the first-stage offered very good precision, so the second-stage effectively extracted the keywords.

In this paper, we propose to use an utterance-level latent topic transition model to estimate the latent topics behind the utterances, and test the performance of such model in extractive speech summarization. In this model, the latent topic weights behind an utterance are estimated, and these topic weights evolve from an utterance to the next in a spoken document based on a topic transition function represented by a matrix. We explore different ways of obtaining such topic transition matrices used in the model, and find using a set of matrices estimated with utterances clustered from a training spoken document set is very useful. This model was shown to be able to offer extra performance improvement when used with the popularly used Probability Latent Semantic Analysis (PLSA) in preliminary experiments on speech summarization.

In this paper, we propose a semantic query expansion approach by extending the query-regularized mixture model to include latent topics and apply it to spoken documents. We also propose to use context feature vectors for spoken segments to train SVM models to enhance the posterior-weighted normalized term frequencies in lattices. Experiments on Mandarin broadcast news showed that this approach offered good improvements when applied on spoken documents including relatively high recognition errors.

In this work, we proposed a new framework for recognition of highly imbalanced code-mixed bilingual speech using an additional frame-level language detector in the conventional recognition system. Blurred posteriorgram features (BPFs) are also proposed to be used in the language detector. The approach was evaluated with real spontaneous lectures offered at National Taiwan University. The highly imbalanced language distribution in code-mixed speech makes the task difficult. Preliminary experimental results showed not only very good performance improvement, but the improvement is complementary to that brought by better acoustic models, whether due to better adaptation approach or increased training data. The code-mixed bilingual speech is frequently used in the daily lives of many people in the globalized world today.

Interaction with user is specially important for spoken con-tent retrieval, not only because of the recognition uncertainty, but because the retrieved spoken content items are difficult to be shown on the screen and difficult to be scanned and selected by the user. The user cannot playback and go through all the retrieved items and then find out they are not what he is look-ing for. In this paper, we propose a new approach for inter-active spoken content retrieval, in which the system can esti-mate the quality of the retrieved results, and take different types of actions to clarify the user’s intention based on an intrinsic policy. The policy is optimized by a Markov Decision Process (MDP) trained with Reinforcement Learning based on a set of pre-defined rewards considering the extra burden given to the user.

Retrieving objects semantically related to the query has been widely studied in text information retrieval. However, when applying the text-based techniques on spoken content, the inevitable recognition errors may seriously degrade the performance. In this paper, we propose to enhance the expected term frequencies estimated from spoken content by acoustic similarity graphs. For each word in the lexicon, a graph is constructed describing acoustic similarity among spoken segments in the archive. Score propagation over the graph helps in estimating the expected term frequencies. The enhanced expected term frequencies can be used in the language modeling retrieval approach, as well as semantic retrieval techniques such as the document expansion based on latent semantic analysis, and query expansion considering both words and latent topic information. Preliminary experiments performed on Mandarin broadcast news indicated that improved performance were achievable under different conditions.

Voice access of cloud applications via smartphones is very attractive today, specifically because a smartphones is used by a single user, so personalized acoustic/language models become feasible. In particular, huge quantities of texts are available within the social networks over the Internet with known authors and given relationships, it is possible to train personalized language models because it is reasonable to assume users with those relationships may share some common subject topics, wording habits and linguistic patterns. In this paper, we propose an adaptation framework for building a robust personalized language model by incorporating the texts the target user and other users had posted on the social networks over the Internet to take care of the linguistic mismatch across different users. Experiments on Facebook dataset showed encouraging improvements in terms of both model perplexity and recognition accuracy with proposed approaches considering relationships among users, similarity based on latent topics, and random walk over a user graph.

Recognition and retrieval are typically viewed as two cascaded independent modules for spoken term detection (STD). Retrieval techniques are assumed to be applied on top of automatic speech recognition (ASR) output, with performance depending on ASR accuracy. We propose a framework that integrates recognition and retrieval and consider them jointly in order to yield better STD performance. This can be achieved either by adjusting the acoustic model parameters (model-based) or by considering detected examples (example-based) using relevance information provided by the user (user relevance feedback) or inferred by the system (pseudo-relevance feedback), either for a given query (short-term context) or by taking into account many previous queries (long-term context). Such relevance feedback approaches have long been used in text information retrieval, but are rarely considered and cannot be directly applied to the retrieval of spoken content. The proposed relevance feedback approaches are specific to spoken content retrieval and are hence very different from those developed for text retrieval, which are applied only to text symbols. We present not only these relevance feedback scenarios and approaches for STD, but also propose a framework to integrate them all together. Preliminary experiments showed significant improvements in each case.

Interaction with users is a powerful strategy that potentially yields better information retrieval for all types of media, including text, images, and videos. While spoken document retrieval (SDR) is a crucial technology for multimedia access in the network era, it is also more challenging than text information retrieval because of the inevitable recognition errors. It is therefore reasonable to consider interactive functionalities for SDR systems. We propose an interactive SDR approach in which given the user's query, the system returns not only the retrieval results but also a short list of key terms describing distinct topics. The user selects these key terms to expand the query if the retrieval results are not satisfactory. The entire retrieval process is organized around a hierarchy of key terms that define the allowable state transitions; this is modeled by a Markov decision process, which is popularly used in spoken dialogue systems. By reinforcement learning with simulated users, the key terms on the short list are properly ranked such that the retrieval success rate is maximized while the number of interactive steps is minimized. Significant improvements over existing approaches were observed in preliminary experiments performed on information needs provided by real users. A prototype system was also implemented.

This paper reports a new approach to improving spoken term detection that uses support vector machine (SVM) with acoustic and linguistic features. As SVM is a good technique for discriminating different features in vector space, we recently proposed to use pseudo-relevance feedback to automatically generate training data for SVM training and use SVM to re-rank the first-pass results considering the context consistency in the lattices. In this paper, we further extend this concept by considering acoustic features at word, phone and HMM state levels and linguistic features of different order. Extensive experiments under various recognition environments demonstrate significant improvements in all cases. In particular, the acoustic features at the HMM state level offered the most significant improvements, and the improvements achieved by acoustic and linguistic features are shown to be additive.

This paper presents a graph-based approach for spoken term detection. Each first-pass retrieved utterance is a node on a graph and the edge between two nodes is weighted by the similarity between the two utterances evaluated in feature space. The score of each node is then modified by the contributions from its neighbors by random walk or its modified version, because utterances similar to more utterances with higher scores should be given higher relevance scores. In this way the global similarity structure of all first-pass retrieved utterances can be jointly considered. Experimental results show that this new approach offers significantly better performance than the previously proposed pseudo-relevance feedback approach, which considers primarily the local similarity relationship between first-pass retrieved utterances, and these two different approaches can be cascaded to provide even better results.

We propose an improved spoken term detection approach that uses support vector machines trained with lattice context consistency. The basic idea is that the same term usually have similar context, while quite different context usually implies the terms are different. Support vector machine can be trained using query context feature vectors obtained from the lattice to estimate better scores for ranking, and significant improvements can be obtained. This process can be performed iteratively and integrated with the pseudo relevance feedback in acoustic feature space proposed previously, both offering further improvements.

Because different indexing features actually have different discriminative capabilities for spoken term detection and different levels of reliability in recognition, it is reasonable to weight the indexing features in the transcribed lattices differently during spoken term detection. In this paper, we present an initial attempt of using two weighting schemes, one context independent (fixed weight for each feature) and one context dependent(different weights for the same feature in different context). These weights can be learned by optimizing a desired spoken term detection performance measure over a training document set and a training query set. Encouraging initial results based on unigrams of Chinese characters and syllables for the corpus of Mandarin broadcast news were obtained from the preliminary experiments.

People usually consider recognition and retrieval as two cascaded independent modules for spoken term detection. Retrieval techniques were assumed to be applied on top of some ASR output, with performance depending on ASR accuracy. In this paper, we propose a new framework: to integrate the two parts into a single task. This can be achieved by adjusting the acoustic model parameters, borrowing the principle of Minimum Classification Error (MCE), based on user feedback. The modified acoustic models then give updated posterior probabilities for the lattice-based structures used in spoken term detection. Encouraging results were obtained on a bilingual course lecture corpus in preliminary experiments.

Abstract In this paper, we propose an improved approach for spokenterm detection using pseudo-relevance feedback. To remedy theproblem of unmatched acoustic models with respect to spokenutterances produced under different acoustic conditions, whichmay give relatively poor recognition output, we integrate therelevance scores derived from the lattices with the DTW dis-tances derived from the feature space of MFCC parametersor phonetic posteriorgrams. These DTW distances are evalu-ated for a carefully selected set of pseudo-relevant utterances,which obtained from the first-pass returned list given by thesearch engine. The utterances on the first-pass returned list arethen reranked accordingly and finally shown to the user. Veryencouraging, performance improvements were obtained in thepreliminary experiments, especially when the acoustic modelsare poorly matched to the spoken utterances.Index Terms: spoken term detection, pseudo-relevance feed-back 1. Introduction Spoken term detection is to return a list of spoken utterancescontaining the term requested by the user. In many approachesof spoken term detection, the spoken utterances are first recog-nized and transformed into transcriptions or lattices by speechrecognition technologies, and then the search engine looksthrough all the transcriptions or lattices very similar to the text-based information retrieval. In this process much of the in-formation in the acoustic signals may be lost in the stage ofspeech recognition, especially when the acoustic models usedare not well matched to the characteristics of the acoustic sig-nals, which naturally results in degraded recognition accuracyand poor detection performance. This is very common in thescenario of spoken term detection, because the huge quantitiesof spoken utterances available over the Internet are naturallyproduced by many different people under many different acous-tic conditions, it is thus very difficult to train a set of acousticmodels well matched to so many different acoustic conditions.As a result, when the relevance scores such as the posteriorprobabilities of the query term derived from transcriptions orlattices are used to rank the retrieved utterances, it is hard tojudge whether a word hypothesis of the query in the transcrip-tions or lattices is a positive target or a false alarm when therecognition output is unreliable. Although many efficient ap-proaches [1, 2, 3] have been proposed to enhance the detectionperformance due to the relatively poor recognition output, thecompensative information straightly from the feature space isnecessary.In text-based information retrieval, even if the texts to beretrieved include all precise words, it is still difficult to retrieveall documents relevant to the query term because many of themdo not include the very short query term entered by the user.However, because many related terms may co-occur in manyrelated documents, a document containing some words appear-ing in some documents identified to be relevant by the searchengine may have high probability to be relevant, even if it doesnot include the query term. For example, a document includingthe words ”George Bush”, ”US”, ”Middle East” may be relevantto a query term of ”White House”, even if it does not includethe query term of ”White House”. In other words, it is possi-ble to retrieve the relevant documents without the query termsince they are ”similar” to some retrieved relevant documentsin some way. Pseudo-relevance feedback, also known as blindrelevance feedback, is one way to realize the above idea. In thisapproach, it is assumed that the set of documents appearing onthe top of the retrieved document list are relevant (or ”pseudo-relevant”), so documents somehow similar to those ”pseudo-relevant” documents can be retrieved, for example, by expand-ing the query with keywords from those ”pseudo-relevant” doc-uments [4]. Similar idea of pseudo-relevance feedback has beenapplied on spoken term detection [5].In this paper, we try to perform similar pseudo-relevancefeedback for spoken term detection as shown in Figure 1. Theupper half of Figure 1 is the conventional spoken term detec-tion. MFCC features were obtained from all spoken utterancesin the archive, speech recognition produces lattices for the ut-terances, and the retrieved engine selects the utterances basedon the relevance scores evaluated from the lattices with respectto the query Qentered by the user. The approach proposedhere in this paper is shown in the lower half of Figure 1. Thefirst-pass returned list is not shown to the user, but instead a”pseudo-relevant utterance set X

This paper presents a new framework integrating different relevance feedback scenarios (pseudo relevance feedback and user relevance feedback in short- and long-term context) and different approaches (model- and example-based) in a spoken term detection system, and shows the retrieval performance can be improved step by step. It is found that short-term context user relevance feedback can further improve the retrieval performance after pseudo relevance feedback, regardless of whether the acoustic models have been adapted by matched data or long-term context user relevance feedback or not. Moreover, model-based and example-based methods are shown to be additive when integrated in short-term context user relevance feedback scenario.

This paper presents the first work known publicly on spoken term detection from bilingual spontaneous speech using code-switched lattice-based structures for word and subword units. The corpus used is the lectures with Chinese as the host language and English as the guest language recorded for a real course offered in National Taiwan University. The techniques reported here have been successfully implemented and tested in a real lecture system now available on-line over the Internet. We also present the approaches of using word fragment as the subword unit for English, and analyse the difficult issues when code-switched lattice-based structures for subword units are used for tasks involving languages of quite different natures.

Lattice-based approaches have been widely used in spoken document retrieval to handle the speech recognition uncertainty and errors. Position Specific Posterior Lattices (PSPL) and Confusion Network (CN) are good examples. It is therefore interesting to derive improved model for spoken document retrieval by properly integrating different versions of lattice-based approaches in order to achieve better performance. In this paper we borrow the framework of dasialearning to rankpsila from text document retrieval and try to integrate it into the scenario of lattice-based spoken document retrieval. Two approaches are considered here, AdaRank and SVM-map. With these approaches, we are able to learn and derived improved models using different versions of PSPL/CN. Preliminary experiments with broadcast news in Mandarin Chinese showed significant improvements.