Although current large audio language models (LALMs) extend text large language models (LLMs) with generic acoustic understanding abilities, they usually suffer from prompt sensitivity, where different instructions of the same intention can yield drastically different outcomes. In this work, we propose AHAMask, where we simply mask some of the attention heads in the decoder-only LLM backbone of LALMs, to trigger specific acoustic task functionalities without instructions. These masks are efficiently obtained by training on an LALM, with the number of trainable parameters equal to the attention head count in its LLM backbone. We show by experiments that applying such selective attention head masks achieves comparable or even better performance than using instructions, either on single or composite tasks. Besides achieving reliable acoustic task specification for LALMs, this also reveals that LALMs exhibit certain ``functional pathways'' in their attention heads.

Flow matching-based generative models offer a principled approach to modeling continuous-time dynamics in speech generation. However, inference is often computationally expensive due to repeated neural network evaluations required by ODE solvers. We propose WaveEx, a training-free and plug-in acceleration framework which replaces portions of ODE integration with wavelet-guided extrapolation. By leveraging the multi-scale structure of latent trajectories, WaveEx predicts future states directly in the frequency domain without additional model evaluations or architectural changes. WaveEx consistently accelerates inference across diverse speech generation tasks. The gains are especially pronounced in tasks like speech synthesis (up to 5.73× speedup) and music generation (2.75×), where flow matching plays a central role in alignment modeling and dense ODE integration. Even in tasks with simpler input-output mappings such as speech enhancement (4.55×) and voice conversion (2.75×), WaveEx still achieves notable acceleration, demonstrating the robustness and generalizability of the approach. These results highlight wavelet-guided extrapolation as a lightweight and broadly applicable alternative to full ODE solving for flow matching-based speech generation.

Wenxi Chen, Ruiqi Yan, Yushen Chen, Zhikang Niu, Ziyang Ma, Xiquan Li, Yuzhe Liang, Wenhanlin, Shunshun Yin, Ming Tao, Xinsheng Wang, Xie Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent years have witnessed remarkable progress in Text-to-Audio Generation (TTA), providing sound creators with powerful tools to transform inspirations into vivid audio.Yet despite these advances, current TTA systems often suffer from slow inference speed, which greatly hinders the efficiency and smoothness of audio creation.In this paper, we present MeanAudio, a fast and faithful textto-audio generator capable of rendering realistic sound with only one function evaluation (1-NFE).MeanAudio leverages: (i) the MeanFlow objective with guided velocity target that significantly accelerates inference speed, (ii) an enhanced Flux-style transformer with dual text encoders for better semantic alignment and synthesis quality, and (iii) an efficient instantaneous-to-mean curriculum that speeds up convergence and enables training on consumer-grade GPUs.Through a comprehensive evaluation study, we demonstrate that MeanAudio achieves state-of-the-art performance in single-step audio generation.Specifically, it achieves a real-time factor (RTF) of 0.013 on a single NVIDIA RTX 3090, yielding a 100x speedup over SOTA diffusion-based TTA systems.Moreover, MeanAudio also shows strong performance in multi-step generation, enabling smooth transitions across successive synthesis steps.

Contrastively pretrained audio–language models (e.g., CLAP) excel at clip-level understanding but struggle with frame-level tasks.Existing extensions fail to exploit the varying granularity of real-world audio–text data, where massive clip-level textual descriptions coexist with limited frame-level annotations. This paper proposes Fine-grained Language-Audio Pretraining (FineLAP), a novel training paradigm that advances both clip- and frame-level alignment in CLAP with heterogeneous data.FineLAP introduces a dual-stream sigmoid loss with a cluster-based sampling strategy to jointly learn from clip- and frame-level supervision. To capture both global semantics and local details, FineLAP uses a decoupled audio projector on top of a self-supervised encoder.To alleviate the scarcity of temporally annotated data, we present FineLAP-100k, a large-scale synthetic SED dataset constructed through a scalable curation pipeline.Extensive experiments demonstrate that FineLAP achieves SOTA performance across multiple audio understanding tasks, including retrieval, classification, sound event detection, and text-to-audio grounding. Ablation studies further show that coarse- and fine-grained alignment are mutually beneficial, providing insights for building better audio-language models (ALMs).

Tianrui Wang, Ziyang Ma, Yizhou Peng, Haoyu Wang, Zhikang Niu, Zikang Huang, Yihao Wu, Yi-Wen Chao, Yu Jiang, Yuheng Lu, Guanrou Yang, Xuanchen Li, Hexin Liu, Chunyu Qiang, Cheng Gong, Yifan Yang, Tianchi Liu, Junyu Wang, Nana Hou, Meng Ge, Fuming You, Yang Wei, Zhongqian Sun, Hu Haifeng, Xiaobao Wang, Eng Siong Chng, Xie Chen, Longbiao Wang, Jianwu Dang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yifan Yang, Bing Han, Hui Wang, Wei Wang, Ziyang Ma, Long Zhou, Zengrui Jin, Guanrou Yang, Tianrui Wang, Xu Tan, Xie Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Chenyuan Zhang, Qiguang Chen, Xie Chen, Zhuotao Tian, Bowen Xing, Meishan Zhang, Libo Qin, Baotian Hu, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The rapid advancement of speech generation technologies in the era of large language models (LLMs) has established discrete speech tokens as a foundational paradigm for speech representation. These tokens, characterized by their discrete, compact, and concise nature, are not only advantageous for efficient transmission and storage, but also inherently compatible with the language modeling framework, enabling seamless integration of speech into text-dominated LLM architectures. Current research categorizes discrete speech tokens into two principal classes: acoustic tokens and semantic tokens, each of which has evolved into a rich research domain characterized by unique design philosophies and methodological approaches. This survey systematically synthesizes the existing taxonomy and recent innovations in discrete speech tokenization, conducts a critical examination of the strengths and limitations of each paradigm, and presents systematic experimental comparisons across token types. Furthermore, we identify persistent challenges in the field and propose potential research directions, aiming to offer actionable insights to inspire future advancements in the development and application of discrete speech tokens.

We present VQTalker, a Vector Quantization-based framework for multilingual talking head generation that addresses the challenges of lip synchronization and natural motion across diverse languages. Our approach is grounded in the phonetic principle that human speech comprises a finite set of distinct sound units (phonemes) and corresponding visual articulations (visemes), which often share commonalities across languages. We introduce a facial motion tokenizer based on Group Residual Finite Scalar Quantization (GRFSQ), which creates a discretized representation of facial features. This method enables comprehensive capture of facial movements while improving generalization to multiple languages, even with limited training data. Building on this quantized representation, we implement a coarse-to-fine motion generation process that progressively refines facial animations. Extensive experiments demonstrate that VQTalker achieves state-of-the-art performance in both video-driven and speech-driven scenarios, particularly in multilingual settings. Notably, our method achieves high-quality results at a resolution of 512 × 512 pixels while maintaining a lower bitrate of approximately 11 kbps. Our work opens new possibilities for cross-lingual talking face generation.

Dialogue serves as the most natural manner of human-computer interaction (HCI). Recent advancements in speech language models (SLM), have significantly enhanced speech-based conversational AI. However, these models are limited to turn-based conversation, lacking the ability to interact with humans in real-time spoken scenarios, for example, being interrupted when the generated content is not satisfactory. To address these limitations, we explore full duplex modeling (FDM) in interactive speech language models (iSLM), focusing on enhancing real-time interaction and, more explicitly, exploring the quintessential ability of interruption. We introduce a novel model design, namely listening-while-speaking language model (LSLM), an end-to-end system equipped with both listening and speaking channels. Our LSLM employs a token-based decoder-only TTS for speech generation and a streaming self-supervised learning (SSL) encoder for real-time audio input. LSLM fuses both channels for autoregressive generation and detects turn-taking in real time. Three fusion strategies—early fusion, middle fusion, and late fusion—are explored, with middle fusion achieving an optimal balance between speech generation and real-time interaction. Two experimental settings, command-based FDM and voice-based FDM, demonstrate LSLM’s robustness to noise and sensitivity to diverse instructions. Our results highlight LSLM’s capability to achieve duplex communication with minimal impact on existing systems. This study aims to advance the development of interactive speech dialogue systems, enhancing their applicability in real-world contexts.

In this paper, we focus on prompting one of the most important tasks in the field of speech processing, i.e., automatic speech recognition (ASR), with speech foundation encoders and large language models (LLM). Despite the growing body of research in this area, we find that many crucial design decisions in LLM-based ASR systems are often inadequately justified. This lack of clarity impedes the field's progress, making it challenging to pinpoint which design choices truly improve model performance. To address these challenges, we conduct a comprehensive series of experiments that explore various aspects, leading to the optimal LLM-based ASR system. We found that delicate designs are not necessary, while a clean setup with little task-specific design is competent. The models achieve strong performance on the Librispeech and Gigaspeech datasets, compared to both LLM-based models and non-LLM-based models. Finally, we explore the capability emergence of LLM-based ASR in the process of modal alignment. We hope that our study can facilitate the research on extending LLM with cross-modality capacity and shed light on the LLM-based ASR community.

The language model (LM) approach based on acoustic and linguistic prompts, such as VALL-E, has achieved remarkable progress in the field of zero-shot audio generation. However, existing methods still have some limitations: 1) repetitions, transpositions, and omissions in the output synthesized speech due to limited alignment constraints between audio and phoneme tokens; 2) challenges of fine-grained control over the synthesized speech with autoregressive (AR) language model; 3) infinite silence generation due to the nature of AR-based decoding, especially under the greedy strategy. To alleviate these issues, we propose ELLA-V, a simple but efficient LM-based zero-shot text-to-speech (TTS) framework, which enables fine-grained control over synthesized audio at the phoneme level. The key to ELLA-V is interleaving sequences of acoustic and phoneme tokens, where phoneme tokens appear ahead of the corresponding acoustic tokens. The experimental findings reveal that our model outperforms baselines in terms of accuracy and delivers more stable results using both greedy and sampling-based decoding strategies.

Yifan Yang, Zheshu Song, Jianheng Zhuo, Mingyu Cui, Jinpeng Li, Bo Yang, Yexing Du, Ziyang Ma, Xunying Liu, Ziyuan Wang, Ke Li, Shuai Fan, Kai Yu, Wei-Qiang Zhang, Guoguo Chen, Xie Chen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

This paper introduces F5-TTS, a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT). Without requiring complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler tokens to the same length as input speech, and then the denoising is performed for speech generation, which was originally proved feasible by E2 TTS. However, the original design of E2 TTS makes it hard to follow due to its slow convergence and low robustness. To address these issues, we first model the input with ConvNeXt to refine the text representation, making it easy to align with the speech. We further propose an inference-time Sway Sampling strategy, which significantly improves our model’s performance and efficiency. This sampling strategy for flow step can be easily applied to existing flow matching based models without retraining. Our design allows faster training and achieves an inference RTF of 0.15, which is greatly improved compared to state-of-the-art diffusion-based TTS models. Trained on a public 100K hours multilingual dataset, our F5-TTS exhibits highly natural and expressive zero-shot ability, seamless code-switching capability, and speed control efficiency. We have released all codes and checkpoints to promote community development, at https://SWivid.github.io/F5-TTS/.

Simultaneous speech translation (SST) outputs translations in parallel with streaming speech input, balancing translation quality and latency.While large language models (LLMs) have been extended to handle the speech modality, streaming remains challenging as speech is prepended as a prompt for the entire generation process.To unlock LLM streaming capability, this paper proposes SimulS2S-LLM, which trains speech LLMs offline and employs a test-time policy to guide simultaneous inference.SimulS2S-LLM alleviates the mismatch between training and inference by extracting boundary-aware speech prompts that allows it to be better matched with text input data.SimulS2S-LLM achieves simultaneous speech-to-speech translation (Simul-S2ST) by predicting discrete output speech tokens and then synthesising output speech using a pretrained vocoder.An incremental beam search is designed to expand the search space of speech token prediction without increasing latency.Experiments on the CVSS speech data show that SimulS2S-LLM offers a better translation quality-latency trade-off than existing methods that use the same training data, such as improving ASR-BLEU scores by 3 points at similar latency.

Yexing Du, Youcheng Pan, Ziyang Ma, Bo Yang, Yifan Yang, Keqi Deng, Xie Chen, Yang Xiang, Ming Liu, Bing Qin. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Audio-Language Models (LALMs) have demonstrated remarkable performance in tasks involving audio perception and understanding, such as speech recognition and audio captioning. However, their reasoning capabilities-critical for solving complex real-world problems-remain underexplored. In this work, we conduct the first exploration into integrating Chain-of-Thought (CoT) reasoning into LALMs to enhance their reasoning ability across auditory modalities. We evaluate representative CoT methods, analyzing their performance in both information extraction and reasoning tasks across sound, music, and speech domains. Our findings reveal that CoT methods significantly improve performance on easy and medium tasks but encounter challenges with hard tasks, where reasoning chains can confuse the model rather than improve accuracy. Additionally, we identify a positive correlation between reasoning path length and accuracy, demonstrating the potential of scaling inference for advanced instruction-following and reasoning. This study not only highlights the promise of CoT in enhancing LALM reasoning capabilities but also identifies key limitations and provides actionable directions for future research. Index Terms-Chain-of-Thought (CoT), Reasoning, Large Audio Language Model (LALM)

Fine-tuning pre-trained speech foundation models for Automatic Speech Recognition (ASR) is prevalent, yet constrained by substantial GPU memory requirements. We introduce ZO-ASR, a memory-efficient Zeroth-Order (ZO) method that avoids Back-Propagation (BP) and activation memory by estimating gradients via forward passes. When combined with SGD optimizer, ZO-ASR-SGD fine-tunes ASR models using only inference memory. Our evaluation spans supervised and unsupervised tasks. For Supervised Domain Adaptation on Whisper-Large-V3, ZO-ASR’s multiple query mechanism enhances robustness and achieves up to an 18.9% relative Word Error Rate reduction over zero-shot baselines, outperforming existing ZO methods. For unsupervised Test-Time Adaptation on Wav2Vec2-Base, ZO-ASR exhibits moderately lower performance compared to first-order optimizer Adam. Our BP-free approach provides a viable solution for fine-tuning ASR models in computationally resource-constrained or gradient-inaccessible scenarios.

Neural speech codecs excel in reconstructing clean speech signals; however, their efficacy in complex acoustic environments and downstream signal processing tasks remains underexplored. In this study, we introduce a novel benchmark named Environment-Resilient Speech Codec Benchmark (ERSB) to systematically evaluate whether neural speech codecs are environment-resilient. Specifically, we assess two key capabilities: (1) robust reconstruction, which measures the preservation of both speech and non-speech acoustic details, and (2) downstream task consistency, which ensures minimal deviation in downstream signal processing tasks when using reconstructed speech instead of the original. Our comprehensive experiments reveal that complex acoustic environments significantly degrade signal reconstruction and downstream task consistency. This work highlights the limitations of current speech codecs and raises a future direction that improves them for greater environmental resilience.

Fine-tuning pre-trained large language models (LLMs) on downstream tasks has achieved significant success across various domains.However, as model sizes grow, traditional firstorder fine-tuning algorithms incur substantial memory overhead due to the need for activation storage for back-propagation (BP).The BP-free Memory-Efficient Zeroth-Order Optimization (MeZO) method estimates gradients through finite differences, avoiding the storage of activation values, and has been demonstrated as a viable approach for fine-tuning large language models.This work proposes the MUltiple-query Memory Efficient Zeroth-Order (MUZO) method, which is based on variance-reduced multiple queries to obtain the average of gradient estimates.When combined with Adam optimizer, MUZO-Adam demonstrates superior performance in fine-tuning various LLMs.Furthermore, we provide theoretical guarantees for the convergence of the MUZO-Adam optimizer.Extensive experiments empirically demonstrate that MUZO-Adam converges better than MeZO-SGD and achieves near first-order optimizer performance on downstream classification, multiple-choice, and generation tasks.

Automated Audio Captioning (AAC) aims to generate natural textual descriptions for input audio signals. Recent progress in audio pre-trained models and large language models (LLMs) has significantly enhanced audio understanding and textual reasoning capabilities, making improvements in AAC possible. In this paper, we propose SLAM-AAC to further enhance AAC with paraphrasing augmentation and CLAP-Refine through LLMs. Our approach uses the self-supervised EAT model to extract fine-grained audio representations, which are then aligned with textual embeddings via lightweight linear layers. The caption generation LLM is efficiently fine-tuned using the LoRA adapter. Drawing inspiration from the back-translation method in machine translation, we implement paraphrasing augmentation to expand the Clotho dataset during pre-training. This strategy helps alleviate the limitation of scarce audio-text pairs and generates more diverse captions from a small set of audio clips. During inference, we introduce the plug-and-play CLAP-Refine strategy to fully exploit multiple decoding outputs, akin to the n-best rescoring strategy in speech recognition. Using the CLAP model for audio-text similarity calculation, we could select the textual descriptions generated by multiple searching beams that best match the input audio. Experimental results show that SLAM-AAC achieves state-of-the-art performance on Clotho V2 and AudioCaps, surpassing previous mainstream models.

Recent TTS models with decoder-only Transformer architecture, such as SPEAR-TTS and VALL-E, achieve impressive naturalness and demonstrate the ability for zero-shot adaptation given a speech prompt. However, such decoder-only TTS models lack monotonic alignment constraints, sometimes leading to hallucination issues such as mispronunciation, word skipping and repeating. To address this limitation, we propose VALL-T, a generative Transducer model that introduces shifting relative position embeddings for input phoneme sequence, explicitly indicating the monotonic generation process while maintaining the architecture of decoder-only Transformer. Consequently, VALL-T retains the capability of prompt-based zero-shot adaptation and demonstrates better robustness against hallucinations with a relative reduction of 28.3% in the word error rate. The audio samples are available at https://cpdu.github.io/vallt.

While automated audio captioning (AAC) has made notable progress, traditional fully supervised AAC models still face two critical challenges: the need for expensive audio-text pair data for training and performance degradation when transferring across domains. To overcome these limitations, we present DRCap, a data-efficient and flexible zero-shot audio captioning system that requires text-only data for training and can quickly adapt to new domains without additional fine-tuning. DRCap integrates a contrastive language-audio pre-training (CLAP) model and a large language model (LLM) as its backbone. During training, the model predicts the ground-truth caption with a fixed text encoder from CLAP, whereas, during inference, the text encoder is replaced with the audio encoder to generate captions for audio clips in a zero-shot manner. To mitigate the modality gap of the CLAP model, we use both the projection strategy from the encoder side and the retrieval-augmented generation strategy from the decoder side. Specifically, audio embeddings are first projected onto a text embedding support to absorb extensive semantic information within the joint multi-modal space of CLAP. At the same time, similar captions retrieved from a datastore are fed as prompts to instruct the LLM, incorporating external knowledge to take full advantage of its strong generative capability. Conditioned on both the projected CLAP embedding and the retrieved similar captions, the model is able to produce a more accurate and semantically rich textual description. By tailoring the text embedding support and the caption datastore to the target domain, DRCap acquires a robust ability to adapt to new domains in a training-free manner. Experimental results demonstrate that DRCap outperforms all other zero-shot models in in-domain scenarios and achieves state-of-the-art performance in cross-domain scenarios.

While automatic speech recognition (ASR) systems have achieved remarkable performance with large-scale datasets, their efficacy remains inadequate in low-resource settings, encompassing dialects, accents, minority languages, and long-tail hotwords, domains with significant practical relevance. With the advent of versatile and powerful text-to-speech (TTS) models, capable of generating speech with human-level naturalness, expressiveness, and diverse speaker profiles, leveraging TTS for ASR data augmentation provides a cost-effective and practical approach to enhancing ASR performance. Comprehensive experiments on an unprecedentedly rich variety of low-resource datasets demonstrate consistent and substantial performance improvements, proving that the proposed method of enhancing low-resource ASR through a versatile TTS model is highly effective and has broad application prospects. Furthermore, we delve deeper into key characteristics of synthesized speech data that contribute to ASR improvement, examining factors such as text diversity, speaker diversity, and the volume of synthesized data, with text diversity being studied for the first time in this work. We hope our findings provide helpful guidance and reference for the practical application of TTS-based data augmentation and push the advancement of low-resource ASR one step further.

We propose a novel and general framework to disentangle video data into its dynamic motion and static content components. Our proposed method is a self-supervised pipeline with less assumptions and inductive biases than previous works: it utilizes a transformer-based architecture to jointly generate flexible implicit features for frame-wise motion and clip-wise content, and incorporates a low-bitrate vector quantization as an information bottleneck to promote disentanglement and form a meaningful discrete motion space. The bitrate-controlled latent motion and content are used as conditional inputs to a denoising diffusion model to facilitate self-supervised representation learning. We validate our disentangled representation learning framework on real-world talking head videos with motion transfer and auto-regressive motion generation tasks. Furthermore, we also show that our method can generalize to other types of video data, such as pixel sprites of 2D cartoon characters. Our work presents a new perspective on self-supervised learning of disentangled video representations, contributing to the broader field of video analysis and generation.

Expressive voice conversion (EVC) aims to modify the speaker identity and emotional style of speech while preserving its content. Existing approaches often focus on disentangling speaker, emotion, and content information but overlook the progressive generation mechanisms in human speech production. To address this, we propose a three-stage framework that includes a speech disentanglement module, a progressive generator, and an acoustic refiner. This framework enables speech pre-trained models to parse linguistic content, emotional style, and speaker identity, which are then progressively integrated into the speech reconstruction branch to generate high-quality speech with replaceable emotional style and speaker identity. Experiments with six different pre-trained models show that our framework activates their disentanglement capabilities, surpassing baseline performance in EVC, and supports speaker and emotion control from different target samples. This framework also provides a valuable reference for evaluating the disentanglement capabilities of speech pre-training models.

Self-supervised learning (SSL) has achieved great success in speech-related tasks. While Transformer and Conformer architectures have dominated SSL backbones, encoders like Zipformer, which excel in automatic speech recognition (ASR), remain unexplored in SSL. Concurrently, inefficiencies in data processing within existing SSL training frameworks, such as fairseq, pose challenges in managing the growing volumes of training data. To address these issues, we propose k2SSL, an open-source framework that offers faster, more memory-efficient, and better-performing self-supervised speech representation learning, focusing on downstream ASR tasks. The optimized HuBERT and proposed Zipformer-based SSL systems exhibit substantial reductions in both training time and memory usage during SSL training. Experiments on LibriSpeech demonstrate that Zipformer Base significantly outperforms HuBERT and WavLM, achieving up to a 34.8% relative WER reduction compared to HuBERT Base after fine-tuning, along with a 3.5x pre-training speedup in GPU hours. When scaled to 60k hours of LibriLight data, Zipformer Large exhibits remarkable efficiency, matching HuBERT Large’s performance while requiring only 5/8 pre-training steps.

MER2025 is the third year of our MER series of challenges. Previously, MER2023 (http://merchallenge.cn/mer2023) focused on multi-label learning, noise robustness, and semi-supervised learning, while MER2024 (https://zeroqiaoba.github.io/MER2024-website) introduced a new track dedicated to open-vocabulary emotion recognition. This year, MER2025 centers on the theme ''When Affective Computing Meets Large Language Models (LLMs)''. We aim to shift the paradigm from traditional categorical frameworks reliant on predefined emotion taxonomies to LLM-driven generative methods, offering innovative solutions for more accurate and reliable emotion understanding. The challenge contains four tracks: MER-SEMI focuses on fixed categorical emotion recognition enhanced by semi-supervised learning; MER-FG explores fine-grained emotions, expanding recognition from basic to nuanced emotional states; MER-DES incorporates multimodal cues (beyond emotion words) into predictions to enhance model interpretability; MER-PR reveals whether emotion prediction results can improve personality recognition performance. For the first three tracks, the baseline code is available at MERTools (https://github.com/zeroQiaoba/MERTools) and datasets can be accessed via Hugging Face (https://huggingface.co/datasets/MERChallenge/MER2025). For the last track, the dataset and baseline code are available on GitHub (https://github.com/cai-cong/MER25_personality).

Recent zero-shot text-to-speech (TTS) systems face a common dilemma: autoregressive (AR) models suffer from slow generation and lack duration controllability, while non-autoregressive (NAR) models lack temporal modeling and typically require complex designs. In this paper, we introduce a novel pseudo-autoregressive (PAR) codec language modeling approach that unifies AR and NAR modeling. Combining explicit temporal modeling from AR with parallel generation from NAR, PAR generates dynamic-length spans at fixed time steps. Building on PAR, we propose PALLE, a two-stage TTS system that leverages PAR for initial generation followed by NAR refinement. In the first stage, PAR progressively generates speech tokens along the time dimension, with each step predicting all positions in parallel but only retaining the left-most span. In the second stage, low-confidence tokens are iteratively refined in parallel, leveraging the global contextual information. Experiments demonstrate that PALLE, trained on LibriTTS, outperforms state-of-the-art systems trained on large-scale data, including F5-TTS, E2-TTS, and MaskGCT, on the LibriSpeech test-clean set in terms of speech quality, speaker similarity, and intelligibility, while achieving up to ten times faster inference speed. Audio samples are available at https://microsoft.com/research/project/vall-e-x/palle.

Human speech goes beyond the mere transfer of information; it is a profound exchange of emotions and a connection between individuals. While Text-to-Speech (TTS) models have made huge progress, they still face challenges in controlling the emotional expression in the generated speech. In this work, we propose EmoVoice, a novel emotion-controllable TTS model that exploits large language models (LLMs) to enable fine-grained freestyle natural language emotion control, and a phoneme boost variant design that makes the model output phoneme tokens and audio tokens in parallel to enhance content consistency, inspired by chain-of-thought (CoT) and modality-of-thought (CoM) techniques. Besides, we introduce EmoVoice-DB, a high-quality 40-hour English emotion dataset featuring expressive speech and fine-grained emotion labels with natural language descriptions. EmoVoice achieves state-of-the-art performance on the English EmoVoice-DB test set using only synthetic training data, and on the Chinese Secap test set using our in-house data. We further investigate the reliability of existing emotion evaluation metrics and their alignment with human perceptual preferences, and explore using SOTA multimodal LLMs GPT-4o-audio and Gemini to assess emotional speech. Dataset, code, checkpoints and demo samples are available at https://github.com/yanghaha0908/EmoVoice.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. These findings underscore the urgent need for greater research attention in audio-language reasoning, including both data and algorithm innovation. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.

While emotional text-to-speech (TTS) has made significant progress, most existing research remains limited to utterance-level emotional expression and fails to support word-level control. Achieving word-level expressive control poses fundamental challenges, primarily due to the complexity of modeling multi-emotion transitions and the scarcity of annotated datasets that capture intra-sentence emotional and prosodic variation. In this paper, we propose WeSCon, the first self-training framework that enables word-level control of both emotion and speaking rate in a pretrained zero-shot TTS model, without relying on datasets containing intra-sentence emotion or speed transitions. Our method introduces a transition-smoothing strategy and a dynamic speed control mechanism to guide the pretrained TTS model in performing word-level expressive synthesis through a multi-round inference process. To further simplify the inference, we incorporate a dynamic emotional attention bias mechanism and fine-tune the model via self-training, thereby activating its ability for word-level expressive control in an end-to-end manner. Experimental results show that WeSCon effectively overcomes data scarcity, achieving state-of-the-art performance in word-level emotional expression control while preserving the strong zero-shot synthesis capabilities of the original TTS model.

The utilization of discrete speech tokens, divided into semantic tokens and acoustic tokens, has been proven superior to traditional acoustic feature mel-spectrograms in terms of naturalness and robustness for text-to-speech (TTS) synthesis. Recent popular models, such as VALL-E and SPEAR-TTS, allow zero-shot speaker adaptation through auto-regressive (AR) continuation of acoustic tokens extracted from a short speech prompt. However, these AR models are restricted to generate speech only in a left-to-right direction, making them unsuitable for speech editing where both preceding and following contexts are provided. Furthermore, these models rely on acoustic tokens, which have audio quality limitations imposed by the performance of audio codec models. In this study, we propose a unified context-aware TTS framework called UniCATS, which is capable of both speech continuation and editing. UniCATS comprises two components, an acoustic model CTX-txt2vec and a vocoder CTX-vec2wav. CTX-txt2vec employs contextual VQ-diffusion to predict semantic tokens from the input text, enabling it to incorporate the semantic context and maintain seamless concatenation with the surrounding context. Following that, CTX-vec2wav utilizes contextual vocoding to convert these semantic tokens into waveforms, taking into consideration the acoustic context. Our experimental results demonstrate that CTX-vec2wav outperforms HifiGAN and AudioLM in terms of speech resynthesis from semantic tokens. Moreover, we show that UniCATS achieves state-of-the-art performance in both speech continuation and editing. Audio samples are available at https://cpdu.github.io/unicats.

Although diffusion models in text-to-speech have become a popular choice due to their strong generative ability, the intrinsic complexity of sampling from diffusion models harms their efficiency. Alternatively, we propose VoiceFlow, an acoustic model that utilizes a rectified flow matching algorithm to achieve high synthesis quality with a limited number of sampling steps. VoiceFlow formulates the process of generating mel-spectrograms into an ordinary differential equation conditional on text inputs, whose vector field is then estimated. The rectified flow technique then effectively straightens its sampling trajectory for efficient synthesis. Subjective and objective evaluations on both single and multi-speaker corpora showed the superior synthesis quality of VoiceFlow compared to the diffusion counterpart. Ablation studies further verified the validity of the rectified flow technique in VoiceFlow.

Zero-shot voice conversion (VC) aims to transfer the source speaker timbre to arbitrary unseen target speaker timbre, while keeping the linguistic content unchanged. Although the voice of generated speech can be controlled by providing the speaker embedding of the target speaker, the speaker similarity still lags behind the ground truth recordings. In this paper, we propose SEF-VC, a speaker embedding free voice conversion model, which is designed to learn and incorporate speaker timbre from reference speech via a powerful position-agnostic cross-attention mechanism, and then reconstruct waveform from HuBERT semantic tokens in a non-autoregressive manner. The concise design of SEF-VC enhances its training stability and voice conversion performance. Objective and subjective evaluations demonstrate the superiority of SEF-VC to generate high-quality speech with better similarity to target reference than strong zero-shot VC baselines, even for very short reference speeches.

While acoustic expressiveness has long been studied in expressive text-to-speech (ETTS), the inherent expressiveness in text lacks sufficient attention, especially for ETTS of artistic works. In this paper, we introduce StoryTTS, a highly ETTS dataset that contains rich expressiveness both in acoustic and textual perspective, from the recording of a Mandarin storytelling show. A systematic and comprehensive labeling framework is proposed for textual expressiveness. We analyze and define speech-related textual expressiveness in StoryTTS to include five distinct dimensions through linguistics, rhetoric, etc. Then we employ large language models and prompt them with a few manual annotation examples for batch annotation. The resulting corpus contains 61 hours of consecutive and highly prosodic speech equipped with accurate text transcriptions and rich textual expressiveness annotations. Therefore, StoryTTS can aid future ETTS research to fully mine the abundant intrinsic textual and acoustic features. Experiments are conducted to validate that TTS models can generate speech with improved expressiveness when integrating with the annotated textual labels in StoryTTS.

In this paper, we explored how to boost speech emotion recognition (SER) with the state-of-the-art speech pre-trained model (PTM), data2vec, text generation technique, GPT-4, and speech synthesis technique, Azure TTS. First, we investigated the representation ability of different speech self-supervised pre-trained models, and we found that data2vec has a good representation ability on the SER task. Second, we employed a powerful large language model (LLM), GPT-4, and emotional text-to-speech (TTS) model, Azure TTS, to generate emotionally congruent text and speech. We carefully designed the text prompt and dataset construction, to obtain the synthetic emotional speech data with high quality. Third, we studied different ways of data augmentation to promote the SER task with synthetic speech, including random mixing, adversarial training, transfer learning, and curriculum learning. Experiments and ablation studies on the IEMOCAP dataset demonstrate the effectiveness of our method, compared with other data augmentation methods, and data augmentation with other synthetic data.

Discrete audio tokens derived from self-supervised learning models have gained widespread usage in speech generation. However, current practice of directly utilizing audio tokens poses challenges for sequence modeling due to the length of the token sequence. Additionally, this approach places the burden on the model to establish correlations between tokens, further complicating the modeling process. To address this issue, we propose acoustic BPE which encodes frequent audio token patterns by utilizing byte-pair encoding. Acoustic BPE effectively reduces the sequence length and leverages the prior morphological information present in token sequence, which alleviates the modeling challenges of token correlation. Through comprehensive investigations on a speech language model trained with acoustic BPE, we confirm the notable advantages it offers, including faster inference and improved syntax capturing capabilities. In addition, we propose a novel rescore method to select the optimal synthetic speech among multiple candidates generated by rich-diversity TTS system. Experiments prove that rescore selection aligns closely with human preference, which highlights acoustic BPE’s potential to other speech generation tasks.

Self-supervised learning (SSL) proficiency in speech-related tasks has driven research into utilizing discrete tokens for speech tasks like recognition and translation, which offer lower storage requirements and great potential to employ natural language processing techniques. However, these studies, mainly single-task focused, faced challenges like overfitting and performance degradation in speech recognition tasks, often at the cost of sacrificing performance in multi-task scenarios. This study presents a comprehensive comparison and optimization of discrete tokens generated by various leading SSL models in speech recognition and synthesis tasks. We aim to explore the universality of speech discrete tokens across multiple speech tasks. Experimental results demonstrate that discrete tokens achieve comparable results against systems trained on FBank features in speech recognition tasks and outperform mel-spectrogram features in speech synthesis in subjective and objective metrics. These findings suggest that universal discrete tokens have enormous potential in various speech-related tasks. Our work is open-source and publicly available at https://github.com/k2-fsa/icefall.

Spatial sound reasoning is a fundamental human skill, enabling us to navigate and interpret our surroundings based on sound. In this paper we present BAT, which combines the spatial sound perception ability of a binaural acoustic scene analysis model with the natural language reasoning capabilities of a large language model (LLM) to replicate this innate ability. To address the lack of existing datasets of in-the-wild spatial sounds, we synthesized a binaural audio dataset using AudioSet and SoundSpaces 2.0. Next, we developed SpatialSoundQA, a spatial sound-based question-answering dataset, offering a range of QA tasks that train BAT in various aspects of spatial sound perception and reasoning. The acoustic front end encoder of BAT is a novel spatial audio encoder named Spatial Audio Spectrogram Transformer, or Spatial-AST, which by itself achieves strong performance across sound event detection, spatial localization, and distance estimation. By integrating Spatial-AST with LLaMA-2 7B model, BAT transcends standard Sound Event Localization and Detection (SELD) tasks, enabling the model to reason about the relationships between the sounds in its environment. Our experiments demonstrate BAT’s superior performance on both spatial sound perception and reasoning, showcasing the immense potential of LLMs in navigating and interpreting complex spatial audio environments.

The paper introduces AniTalker, an innovative framework designed to generate lifelike talking faces from a single portrait. Unlike existing models that primarily focus on verbal cues such as lip synchronization and fail to capture the complex dynamics of facial expressions and nonverbal cues, AniTalker employs a universal motion representation. This innovative representation effectively captures a wide range of facial dynamics, including subtle expressions and head movements. AniTalker enhances motion depiction through two self-supervised learning strategies: the first involves reconstructing target video frames from source frames within the same identity to learn subtle motion representations, and the second develops an identity encoder using metric learning while actively minimizing mutual information between the identity and motion encoders. This approach ensures that the motion representation is dynamic and devoid of identity-specific details, significantly reducing the need for labeled data. Additionally, the integration of a diffusion model with a variance adapter allows for the generation of diverse and controllable facial animations. This method not only demonstrates AniTalker's capability to create detailed and realistic facial movements but also underscores its potential in crafting dynamic avatars for real-world applications. Synthetic results can be viewed at https://github.com/X-LANCE/AniTalker.

Built upon vector quantization (VQ), discrete audio codec models have achieved great success in audio compression and autoregressive audio generation. However, existing models face substantial challenges in perceptual quality and signal distortion, especially when operating in extremely low bandwidth, rooted in the sensitivity of the VQ codebook to noise. This degradation poses significant challenges for several downstream tasks, such as codec-based speech synthesis. To address this issue, we propose a novel VQ method, Normal Distribution-based Vector Quantization (NDVQ), by introducing an explicit margin between the VQ codes via learning a variance. Specifically, our approach involves mapping the waveform to a latent space and quantizing it by selecting the most likely normal distribution, with each codebook entry representing a unique normal distribution defined by its mean and variance. Using these distribution-based VQ codec codes, a decoder reconstructs the input waveform. NDVQ is trained with additional distribution-related losses, alongside reconstruction and discrimination losses. Experiments demonstrate that NDVQ outperforms existing audio compression baselines, such as EnCodec, in terms of audio quality and zero-shot TTS, particularly in very low bandwidth scenarios.

Recent popular decoder-only text-to-speech models are known for their ability of generating natural-sounding speech. However, such models sometimes suffer from word skipping and repeating due to the lack of explicit monotonic alignment constraints. In this paper, we notice from the attention maps that some particular attention heads of the decoder-only model indicate the alignments between speech and text. We call the attention maps of those heads Alignment-Emerged Attention Maps (AEAMs). Based on this discovery, we propose a novel inference method without altering the training process, named Attention-Constrained Inference (ACI), to facilitate monotonic synthesis. It first identifies AEAMs using the Attention Sweeping algorithm and then applies constraining masks on AEAMs. Our experimental results on decoderonly TTS model VALL-E show that the WER of synthesized speech is reduced by up to 20.5% relatively with ACI while the naturalness and speaker similarity are comparable.

Contextual ASR or hotword customization holds substantial practical value. Despite the impressive performance of current end-to-end (E2E) automatic speech recognition (ASR) systems, they often face challenges in accurately recognizing rare words. Typical E2E contextual ASR models commonly feature complex architectures and decoding mechanisms, limited in performance and susceptible to interference from distractor words. With large language model (LLM)-based ASR models emerging as the new mainstream, we propose a CTC-Assisted LLM-Based Contextual ASR model with an efficient filtering algorithm. By using coarse CTC decoding results to filter potential relevant hotwords and incorporating them into LLM prompt input, our model attains WER/B-WER of $1.27 \% / 3.67 \%$ and $2.72 \% / 8.02 \%$ on the Librispeech test-clean and test-other sets targeting on recognizing rare long-tail words, demonstrating significant improvements compared to the baseline LLM-based ASR model, and substantially surpassing other related work. More remarkably, with the help of the large language model and proposed filtering algorithm, our contextual ASR model still performs well with 2000 biasing words.1

Long-form automatic speech recognition (ASR) has obtained increasing interest in recent years, as it captures the relationship among consecutive historical sentences while decoding the current sentence. In this paper, we propose two novel approaches, which integrate long-form information into the factorized neural transducer (FNT) based architecture in both non-streaming (referred to asLongFNT) and streaming (referred to asSLongFNT) scenarios. We first investigate whether long-form transcriptions can improve the vanilla conformer transducer (C-T) models. Our experiments indicate that the vanilla C-T models do not exhibit improved performance when utilizing long-form transcriptions, possibly due to the predictor network of C-T models not functioning as a pure language model. Instead, FNT shows its potential in utilizing long-form information, where we propose theLongFNTmodel and explore the impact of long-form information in both text (LongFNT-Text) and speech (LongFNT-Speech). The proposed LongFNT-Text and LongFNT-Speech models further complement each other to achieve better performance, with transcription history proving more valuable to the model. The effectiveness of our LongFNT approach is evaluated on LibriSpeech and GigaSpeech corpora, and obtains relative 19% and 12% word error rate reduction, respectively. Furthermore, we extend the LongFNT model to the streaming scenario, which is namedSLongFNT, consisting of SLongFNT-Text and SLongFNT-Speech approaches to utilize long-form text and speech information. Experiments show that the proposed SLongFNT model achieves relative 26% and 17% WER reduction on LibriSpeech and GigaSpeech respectively while keeping a good latency, compared to the FNT baseline. Overall, our proposedLongFNTandSLongFNThighlight the significance of considering long-form speech and transcription knowledge for improving both non-streaming and streaming speech recognition systems.

Text-based speech editing aims at manipulating part of real audio by modifying the corresponding transcribed text, without being discernible by human auditory system. With the enhanced capability of neural Text-to-speech (TTS), researchers try to tackle speech editing problems with TTS methods. In this paper, we propose E$^{3}$TTS, a.k.a. end-to-end text-based speech editing TTS system, which combines a text encoder, a speech encoder, and a joint net for speech synthesis and speech editing. E$^{3}$TTS can insert, replace, and delete speech content at will, by manipulating the given text. Experiments show that our speech editing outperforms strong baselines on HiFiTTS and LibriTTS datasets, speakers of which are seen or unseen, respectively. Further, we introduce E$^{3}$TTS into data augmentation for automatic speech recognition (ASR) to mitigate the data insufficiency problem in code-switching and named entity recognition scenarios1. E$^{3}$TTS retains the coherence and reality of the recorded audio compared to past data augmentation methods. The experimental results show significant performance improvements over baseline systems with traditional TTS-based data augmentation. The code and samples of the proposed speech editing model are available at this repository.2

Self-supervised learning (SSL) has achieved great success in speech processing, but always with a large model size to increase the modeling capacity. This may limit its potential applications due to the expensive computation and memory costs introduced by the oversize model. Compression for SSL models has become an important research direction of practical value. To this end, we explore the effective distillation of HuBERT-based SSL models for automatic speech recognition. First, a comprehensive study of different student model structures is conducted. On top of this, as a supplement to the regression loss widely adopted in previous works, a discriminative loss is introduced for HuBERT to enhance the distillation performance, especially in low-resource scenarios. In addition, we design a simple and effective algorithm to distill the front-end input from waveform to Fbank feature, resulting in 17% parameter reduction and doubling inference speed, at marginal performance degradation.

Recent years have witnessed significant advancements in self-supervised learning (SSL) methods for speech-processing tasks. Various speech-based SSL models have been developed and present promising performance on a range of downstream tasks including speech recognition. However, existing speech-based SSL models face a common dilemma in terms of computational cost, which might hinder their potential application and in-depth academic research. To address this issue, we first analyze the computational cost of different modules during HuBERT pre-training and then introduce a stack of efficiency optimizations, which is named Fast-HuBERT in this paper. The proposed Fast-HuBERT can be trained in 1.1 days with 8 V100 GPUs on the Librispeech 960 h benchmark, without performance degradation, resulting in a 5.2x speedup, compared to the original implementation. Moreover, we explore two well-studied techniques in the Fast-HuBERT and demonstrate consistent improvements as reported in previous work.11The code for Fast-HuBERT training is available at https://github.com/yanghaha0908/FastHuBERT

Audio-driven talking face has attracted broad interest from academia and industry recently. However, data acquisition and labeling in audio-driven talking face are labor-intensive and costly. The lack of data resource results in poor synthesis effect. To alleviate this issue, we propose to use TTS (Text-To-Speech) for data augmentation to improve few-shot ability of the talking face system. The misalignment problem brought by the TTS audio is solved with the introduction of soft-DTW, which is first adopted in the talking face task. Moreover, features extracted by HuBERT are explored to utilize underlying information of audio, and found to be superior over other features. The proposed method achieves 17%, 14%, 38% dominance on MSE score, DTW score and user study preference repectively over the baseline model, which shows the effectiveness of improving few-shot learning for talking face system with TTS augmentation.

Recent years have witnessed a boom in self-supervised learning (SSL) in various areas including speech processing. Speech based SSL models present promising performance in a range of speech related tasks. However, the training of SSL models is computationally expensive and a common practice is to fine-tune a released SSL model on the specific task. It is essential to use consistent front-end input during pre-training and fine-tuning. This consistency may introduce potential issues when the optimal front-end is not the same as that used in pre-training. In this paper, we propose a simple but effective front-end adapter to address this front-end discrepancy. By minimizing the distance between the outputs of different front-ends, the filterbank feature (Fbank) can be compatible with SSL models which are pre-trained with waveform. The experiment results demonstrate the effectiveness of our proposed front-end adapter on several popular SSL models for the speech recognition task.

Traditional automatic speech recognition (ASR) systems usually focus on individual utterances, without considering long-form speech with useful historical information, which is more practical in real scenarios. Simply attending longer transcription history for a vanilla neural transducer model shows no much gain in our preliminary experiments, since the prediction network is not a pure language model. This motivates us to leverage the factorized neural transducer structure, containing a real language model, the vocabulary predictor. We propose the LongFNT-Text architecture, which fuses the sentence-level long-form features directly with the output of the vocabulary predictor and then embeds token-level long-form features inside the vocabulary predictor, with a pre-trained contextual encoder RoBERTa to further boost the performance. Moreover, we propose the LongFNT architecture by extending the long-form speech to the original speech input and achieve the best performance. The effectiveness of our LongFNT approach is validated on LibriSpeech and GigaSpeech corpora with 19% and 12% relative word error rate (WER) reduction, respectively.

End-to-end automatic speech recognition (ASR) systems have gained popularity given their simplified architecture and promising results. However, text-only domain adaptation remains a big challenge for E2E systems. Text-to-speech (TTS) based approaches fine-tune ASR models by synthesized speech with an auxiliary TTS model, thus increase deployment costs. Language model (LM) fusion based approaches can achieve good performance but are sensitive to interpolation parameters. In order to factorize out the language component in the AED model, we propose the factorized attention-based encoder-decoder (Factorized AED) model whose decoder takes as input the posterior probabilities of a jointly trained LM. Moreover, in the context of domain adaptation, the domain specific LM serves as a plug-and-play component for a well-trained factorized AED model. In-domain experiments on LibriSpeech and out-of-domain experiments adapting from LibriSpeech to a variety of domains in GigaSpeech are conducted to validate the effectiveness of our proposed methods. Results show 20% / 24% relative word error rate (WER) reduction for LibriSpeech test sets and 8 ∼34% relative WER reduction for 8 GigaSpeech target domains test sets compared to the AED baseline.

Although current neural text-to-speech (TTS) models are able to generate high-quality speech, intensity controllable emotional TTS is still a challenging task. Most existing methods need external optimizations for intensity calculation, leading to suboptimal results or degraded quality. In this paper, we propose EmoDiff, a diffusion-based TTS model where emotion intensity can be manipulated by a proposed soft-label guidance technique derived from classifier guidance. Specifically, instead of being guided with a one-hot vector for the specified emotion, EmoDiff is guided with a soft label where the value of the specified emotion and Neutral is set to α and 1 − α respectively. The α here represents the emotion intensity and can be chosen from 0 to 1. Our experiments show that EmoDiff can precisely control the emotion intensity while maintaining high voice quality. Moreover, diverse speech with specified emotion intensity can be generated by sampling in the reverse denoising process.

In recent years, self-supervised learning (SSL) has achieved tremendous success in various speech tasks due to its power to extract representations from massive unlabeled data. However, compared with tasks such as speech recognition (ASR), the improvements from SSL representation in speech separation (SS) and enhancement (SE) are considerably smaller. Based on HuBERT, this work investigates improving the SSL model for SS and SE. We first update HuBERT’s masked speech prediction (MSP) objective by integrating the separation and denoising terms, resulting in a multiple pseudo label pre-training scheme, which significantly improves HuBERT’s performance on SS and SE but degrades the performance on ASR. To maintain its performance gain on ASR, we further propose an adapter-based architecture for HuBERT’s Transformer encoder, where only a few parameters of each layer are adjusted to the multiple pseudo label MSP while other parameters remain frozen as default HuBERT. Experimental results show that our proposed adapter-based multiple pseudo label HuBERT yield consistent and significant performance improvements on SE, SS, and ASR tasks, with a faster pretraining speed, at only marginal parameters increase.

While recent research has made significant progress in speech-driven talking face generation, the quality of the generated video still lags behind that of real recordings. One reason for this is the use of handcrafted intermediate representations like facial landmarks and 3DMM coefficients, which are designed based on human knowledge and are insufficient to precisely describe facial movements. Additionally, these methods require an external pretrained model for extracting these representations, whose performance sets an upper bound on talking face generation. To address these limitations, we propose a novel method called DAE-Talker that leverages data-driven latent representations obtained from a diffusion autoencoder (DAE). DAE contains an image encoder that encodes an image into a latent vector and a DDIM-based image decoder that reconstructs the image from it. We train our DAE on talking face video frames and then extract their latent representations as the training target for a Conformer-based speech2latent model. During inference, DAE-Talker first predicts the latents from speech and then generates the video frames with the image decoder in DAE from the predicted latents. This allows DAE-Talker to synthesize full video frames and produce natural head movements that align with the content of speech, rather than relying on a predetermined head pose from a template video. We also introduce pose modelling in speech2latent for pose controllability. Additionally, we propose a novel method for generating continuous video frames with the DDIM-based image decoder trained on individual frames, eliminating the need for modelling the joint distribution of consecutive frames directly. Our experiments show that DAE-Talker outperforms existing popular methods in lip-sync, video fidelity, and pose naturalness. We also conduct ablation studies to analyze the effectiveness of the proposed techniques and demonstrate the pose controllability of DAE-Talker.

Achieving high fidelity and speaker similarity in text-to-speech speaker adaptation with limited amount of data is a challenging task. Most existing methods only consider adapting to the timbre of the target speakers but fail to capture their speaking styles from little data. In this work, we propose a novel TTS system, TN-VQTTS, which leverages timbre-normalized vector-quantized (TN-VQ) acoustic feature for speaker adaptation with little data. With the TN-VQ feature, speaking style and timbre can be effectively decomposed and controlled by the acoustic model and the vocoder separately of VQTTS. Such decomposition enables us to finely mimic both the two characteristics of the target speaker in adaptation with little data. Specifically, we first reduce the dimensionality of self-supervised VQ acoustic feature via PCA and normalize its timbre with a normalizing flow model. The feature is then quantized with k-means and used as the TN-VQ feature for a multi-speaker VQ-TTS system. Furthermore, we optimize timbre-independent style embeddings of the training speakers jointly with the acoustic model and store them in a lookup table. The embedding table later serves as a selectable codebook or a group of basis for representing the style of unseen speakers. Our experiments on LibriTTS dataset first show that the proposed model architecture for VQ feature achieves better performance in multi-speaker text-to-speech synthesis than several existing methods. We also find that the reconstruction performance and the naturalness are almost unchanged after applying timbre normalization and k-means quantization. Finally, we show that TN-VQTTS achieves better performance on speaker similarity in adaptation than both speaker embedding based adaptation method and fine-tuning based baseline AdaSpeech.

In recent years, end-to-end (E2E) based automatic speech recognition (ASR) systems have achieved great success due to their simplicity and promising performance. Neural Transducer based models are increasingly popular in streaming E2E based ASR systems and have been reported to outperform the traditional hybrid system in some scenarios. However, the joint optimization of acoustic model, lexicon and language model (LM) in neural Transducer also brings about challenges in adapting ASR using just adaptation text. This draw-back might prevent their potential applications in practice. In order to address this issue, we propose a novel model, factorized neural Transducer, by factorizing the blank and vocabulary prediction, and adopting a standalone language model for the vocabulary prediction. It is expected that this factorization can transfer the improvement of the standalone language model to the Transducer for speech recognition, which allows various language model adaptation techniques to be applied. We demonstrate that the proposed factorized neural Transducer yields 15.4% to 19.4% WER improvements when out-of-domain text data is used for language model adaptation, at the cost of a minor degradation in WER on a general test set.

The mainstream neural text-to-speech(TTS) pipeline is a cascade system, including an acoustic model(AM) that predicts acoustic feature from the input transcript and a vocoder that generates waveform according to the given acoustic feature. However, the acoustic feature in current TTS systems is typically mel-spectrogram, which is highly correlated along both time and frequency axes in a complicated way, leading to a great difficulty for the AM to predict. Although high-fidelity audio can be generated by recent neural vocoders from ground-truth(GT) mel-spectrogram, the gap between the GT and the predicted mel-spectrogram from AM degrades the performance of the entire TTS system. In this work, we propose VQTTS, consisting of an AM txt2vec and a vocoder vec2wav, which uses self-supervised vector-quantized(VQ) acoustic feature rather than mel-spectrogram. We redesign both the AM and the vocoder accordingly. In particular, txt2vec basically becomes a classification model instead of a traditional regression model while vec2wav uses an additional feature encoder before HifiGAN generator for smoothing the discontinuous quantized feature. Our experiments show that vec2wav achieves better reconstruction performance than HifiGAN when using self-supervised VQ acoustic feature. Moreover, our entire TTS system VQTTS achieves state-of-the-art performance in terms of naturalness among all current publicly available TTS systems.

Text-only adaptation of an end-to-end (E2E) model remains a challenging task for automatic speech recognition (ASR).Language model (LM) fusion-based approaches require an additional external LM during inference, significantly increasing the computation cost.To overcome this, we propose an internal LM adaptation (ILMA) of the E2E model using text-only data.Trained with audio-transcript pairs, an E2E model implicitly learns an internal LM that characterizes the token sequence probability which is approximated by the E2E model output after zeroing out the encoder contribution.During ILMA, we fine-tune the internal LM, i.e., the E2E components excluding the encoder, to minimize a cross-entropy loss.To make ILMA effective, it is essential to train the E2E model with an internal LM loss besides the standard E2E loss.Furthermore, we propose to regularize ILMA by minimizing the Kullback-Leibler divergence between the output distributions of the adapted and unadapted internal LMs.ILMA is the most effective when we update only the last linear layer of the joint network.ILMA enables a fast text-only adaptation of the E2E model without increasing the run-time computational cost.Experimented with 30K-hour trained transformer transducer models, ILMA achieves up to 34.9% relative word error rate reduction from the unadapted baseline.

Recently, Transformer based end-to-end models have achieved great success in many areas including speech recognition. However, compared to LSTM models, the heavy computational cost of the Transformer during inference is a key issue to prevent their applications. In this work, we explored the potential of Transformer Transducer (T-T) models for the fist pass decoding with low latency and fast speed on a large-scale dataset. We combine the idea of Transformer- XL and chunk-wise streaming processing to design a streamable Transformer Transducer model. We demonstrate that T-T outperforms the hybrid model, RNN Transducer (RNN-T), and streamable Transformer attention-based encoder-decoder model in the streaming scenario. Furthermore, the runtime cost and latency can be optimized with a relatively small look-ahead.

The efficacy of external language model (LM) integration with existing end-to-end (E2E) automatic speech recognition (ASR) systems can be improved significantly using the internal language model estimation (ILME) method [1]. In this method, the internal LM score is subtracted from the score obtained by interpolating the E2E score with the external LM score, during inference. To improve the ILME-based inference, we propose an internal LM training (ILMT) method to minimize an additional internal LM loss by updating only the E2E model components that affect the internal LM estimation. ILMT encourages the E2E model to form a standalone LM inside its existing components, without sacrificing ASR accuracy. After ILMT, the more modular E2E model with matched training and inference criteria enables a more thorough elimination of the source-domain internal LM, and therefore leads to a more effective integration of the target-domain external LM. Experimented with 30K-hour trained recurrent neural network transducer and attention-based encoder- decoder models, ILMT with ILME-based inference achieves up to 31.5% and 11.4% relative word error rate reductions from standard E2E training with Shallow Fusion on out-of-domain LibriSpeech and in-domain Microsoft production test sets, respectively.

Many state-of-the-art ML results have been obtained by scaling up the number of parameters in existing models. However, parameters and activations for such large models often do not fit in the memory of a single accelerator device; this means that it is necessary to distribute training of large models over multiple accelerators. In this work, we propose PipeDream-2BW, a system that supports memory-efficient pipeline parallelism. PipeDream-2BW uses a novel pipelining and weight gradient coalescing strategy, combined with the double buffering of weights, to ensure high throughput, low memory footprint, and weight update semantics similar to data parallelism. In addition, PipeDream-2BW automatically partitions the model over the available hardware resources, while respecting hardware constraints such as memory capacities of accelerators and interconnect topologies. PipeDream-2BW can accelerate the training of large GPT and BERT language models by up to 20x with similar final model accuracy.

Integrating external language models (LMs) into end-to-end (E2E) models remains a challenging task for domain-adaptive speech recognition.Recently, internal language model estimation (ILME)-based LM fusion has shown significant word error rate (WER) reduction from Shallow Fusion by subtracting a weighted internal LM score from an interpolation of E2E model and external LM scores during beam search.However, on different test sets, the optimal LM interpolation weights vary over a wide range and have to be tuned extensively on well-matched validation sets.In this work, we perform LM fusion in the minimum WER (MWER) training of an E2E model to obviate the need for LM weights tuning during inference.Besides MWER training with Shallow Fusion (MWER-SF), we propose a novel MWER training with ILME (MWER-ILME) where the ILME-based fusion is conducted to generate N-best hypotheses and their posteriors.Additional gradient is induced when internal LM is engaged in MWER-ILME loss computation.During inference, LM weights pre-determined in MWER training enable robust LM integrations on test sets from different domains.Experimented with 30K-hour trained transformer transducers, MWER-ILME achieves on average 8.8% and 5.8% relative WER reductions from MWER and MWER-SF training, respectively, on 6 different test sets.

The external language models (LM) integration remains a challenging task for end-to-end (E2E) automatic speech recognition (ASR) which has no clear division between acoustic and language models. In this work, we propose an internal LM estimation (ILME) method to facilitate a more effective integration of the external LM with all pre-existing E2E models with no additional model training, including the most popular recurrent neural network transducer (RNN-T) and attention-based encoder-decoder (AED) models. Trained with audio-transcript pairs, an E2E model implicitly learns an internal LM that characterizes the training data in the source domain. With ILME, the internal LM scores of an E2E model are estimated and subtracted from the log-linear interpolation between the scores of the E2E model and the external LM. The internal LM scores are approximated as the output of an E2E model when eliminating its acoustic components. ILME can alleviate the domain mismatch between training and testing, or improve the multi-domain E2E ASR. Experimented with 30K-hour trained RNN-T and AED models, ILME achieves up to 15.5% and 6.8% relative word error rate reductions from Shallow Fusion on out-of-domain LibriSpeech and in-domain Microsoft production test sets, respectively.

The high memory consumption and computational costs of Recurrent neural network language models (RNNLMs) limit their wider application on resource constrained devices. In recent years, neural network quantization techniques that are capable of producing extremely low-bit compression, for example, binarized RNNLMs, are gaining increasing research interests. Directly training of quantized neural networks is difficult. By formulating quantized RNNLMs training as an optimization problem, this paper presents a novel method to train quantized RNNLMs from scratch using alternating direction methods of multipliers (ADMM). This method can also flexibly adjust the trade-off between the compression rate and model performance using tied low-bit quantization tables. Experiments on two tasks: Penn Treebank (PTB), and Switchboard (SWBD) suggest the proposed ADMM quantization achieved a model size compression factor of up to 31 times over the full precision baseline RNNLMs. Faster convergence of 5 times in model training over the baseline binarized RNNLM quantization was also obtained.

Maximum entropy language models (MaxEnt LMs) are log-linear models which are able to incorporate various hand-crafted features and non-linguistic information. Standard MaxEnt LMs are computationally heavy for tasks with a large vocabulary size due to the expensive normalization computation in the denominator. To address this issue, most recent works on MaxEnt LMs have used class based MaxEnt LMs. However, the performance of class based MaxEnt LMs might be sensitive to word clustering and it is also time-consuming to generate high-quality word classes. Motivated by the recent success of sampling techniques in accelerating the training of neural network language models, in this paper, three widely used sampling techniques, importance sampling, noise contrastive estimation (NCE) and sampled softmax, are investigated for the MaxEnt LM training. Experimental results on the Google One Billion corpus and an internal speech recognition system demonstrate the effectiveness of sampled softmax and NCE on MaxEnt LM training. However, importance sampling is not effective for MaxEnt LM training despite its similarity to sampled softmax. To our knowledge, this is the first work applying sampling techniques on MaxEnt LM training.

Recurrent neural network language models (RNNLMs) have shown superior performance across a range of speech recognition tasks. At the heart of all RNNLMs, the activation functions play a vital role to control the information flows and tracking longer history contexts that are useful for predicting the following words. Long short-term memory (LSTM) units are well known for such ability and thus widely used in current RNNLMs. However, the deterministic parameter estimates in LSTM RNNLMs are prone to over-fitting and poor generalization when given limited training data. Furthermore, the precise forms of activations in LSTM have been largely empirically set for all cells at a global level. In order to address these issues, this paper introduces Gaussian process (GP) LSTM RNNLMs. In addition to modeling parameter uncertainty under a Bayesian framework, it also allows the optimal forms of gates being automatically learned for individual LSTM cells. Experiments were conducted on three tasks: the Penn Treebank (PTB) corpus, Switchboard conversational telephone speech (SWBD) and the AMI meeting room data. The proposed GP-LSTM RNNLMs consistently outperform the baseline LSTM RNNLMs in terms of both perplexity and word error rate.

Recurrent neural network language models (RNNLMs) have become an increasing popular choice for state-of-the-art speech recognition systems. RNNLMs are normally trained by minimizing the cross entropy (CE) using the stochastic gradient descent (SGD) algorithm. However, the SGD method doesn't consider the correlation between parameters and therefore can lead to unstable and slow convergence in training. Second-order optimization methods provide a possible solution to this issue. However these methods are either computationally heavy or do not have competitive performance. In this paper, a novel optimization method - stochastic natural gradient based on minimum variance assumption (SNGM) is proposed for training RNNLMs. It allows the natural gradient method to operate at a comparable training efficiency to the SGD method. By modifying the gradient according to the local curvature of the KL-divergence between current and updated probabilistic distributions, the proposed SNGM approach is shown to outperform both the SGD and limited memory BFGS methods across three tasks: Penn Treebank, Switchboard conversational speech recognition and AMI meeting room transcription in terms of both perplexity and word error rate.

Language modeling is a crucial component in a wide range of applications including speech recognition. Language models (LMs) are usually constructed by splitting a sentence into words and computing the probability of a word based on its word history. This sentence probability calculation, making use of conditional probability distributions, assumes that there is little impact from approximations used in the LMs, including the word history representations and finite training data. This motivates examining models that make use of additional information from the sentence. In this paper, future word information, in addition to the history, is used to predict the probability of the current word. For recurrent neural network LMs (RNNLMs), this information can be encapsulated in a bi-directional model. However, if used directly, this form of model is computationally expensive when trained on large quantities of data, and can be problematic when used with word lattices. This paper proposes a novel neural network language model structure, the succeeding-word RNNLM, su-RNNLM, to address these issues. Instead of using a recurrent unit to capture the complete future word contexts, a feedforward unit is used to model a fixed finite number of succeeding words. This is more efficient in training than bi-directional models and can be applied to lattice rescoring. The generated lattices can be used for downstream applications, such as confusion network decoding and keyword search. Experimental results on speech recognition and keyword spotting tasks illustrate the empirical usefulness of future word information, and the flexibility of the proposed model to represent this information.

Recurrent neural network language models (RNNLM) have become an increasingly popular choice for state-of-the-art speech recognition systems. RNNLMs are normally trained by minimizing the cross entropy (CE) using the stochastic gradient descent (SGD) algorithm. The SGD method only uses first-order derivatives and no higher order gradient information is used to consider the correlation between model parameters. It is unable to fully capture the curvature of the error cost function. This can lead to slow convergence in model training. In this paper, a limited-memory Broyden Fletcher Goldfarb Shannon (L-BFGS) based second order optimization technique is proposed for RNNLMs. This method efficiently approximates the matrix-vector product between the inverse Hessian and gradient vector via a recursion over past gradients with a compact memory requirement. Consistent perplexity and error rate reductions are obtained over the SGD method on two speech recognition tasks: Switchboard English and Babel Cantonese. A faster convergence and speed up in RNNLM training time was also obtained.

State-of-the-art English automatic speech recognition systems typically use phonetic rather than graphemic lexicons. Graphemic systems are known to perform less well for English as the mapping from the written form to the spoken form is complicated. However, in recent years the representational power of deep-learning based acoustic models has improved, raising interest in graphemic acoustic models for English, due to the simplicity of generating the lexicon. In this paper, phonetic and graphemic models are compared for an English Multi-Genre Broadcast transcription task. A range of acoustic models based on lattice-free MMI training are constructed using phonetic and graphemic lexicons. For this task, it is found that having a long-span temporal history reduces the difference in performance between the two forms of models. In addition, system combination is examined, using parameter smoothing and hypothesis combination. As the combination approaches become more complicated the difference between the phonetic and graphemic systems further decreases. Finally, for all configurations examined the combination of phonetic and graphemic systems yields consistent gains.

In this paper we describe an extension of the Kaldi software toolkit to support neural-based language modeling, intended for use in automatic speech recognition (ASR) and related tasks. We combine the use of subword features (letter n-grams) and one-hot encoding of frequent words so that the models can handle large vocabularies containing infrequent words. We propose a new objective function that allows for training of unnormalized probabilities. An importance sampling based method is supported to speed up training when the vocabulary is large. Experimental results on five corpora show that Kaldi-RNNLM rivals other recurrent neural network language model toolkits both on performance and training speed.

Making predictions of the following word given the back history of words may be challenging without meta-information such as the topic. Standard neural network language models have an implicit representation of the topic via the back history of words. In this work a more explicit form of topic representation is used via an attention mechanism. Though this makes use of the same information as the standard model, it allows parameters of the network to focus on different aspects of the task. The attention model provides a form of topic representation that is automatically learned from the data. Whereas the recurrent model deals with the (conditional) history representation. The combined model is expected to reduce the stress on the standard model to handle multiple aspects. Experiments were conducted on the Penn Tree Bank and BBC Multi-Genre Broadcast News (MGB) corpora, where the proposed approach outperforms standard forms of recurrent models in perplexity. Finally, N-best list rescoring for speech recognition in the MGB3 task shows word error rate improvements over comparable standard form of recurrent models.

Recently, bidirectional recurrent network language models (bi-RNNLMs) have been shown to outperform standard, unidirectional, recurrent neural network language models (uni-RNNLMs) on a range of speech recognition tasks. This indicates that future word context information beyond the word history can be useful. However, bi-RNNLMs pose a number of challenges as they make use of the complete previous and future word context information. This impacts both training efficiency and their use within a lattice rescoring framework. In this paper these issues are addressed by proposing a novel neural network structure, succeeding word RNNLMs (suRNNLMs). Instead of using a recurrent unit to capture the complete future word contexts, a feedforward unit is used to model a finite number of succeeding, future, words. This model can be trained much more efficiently than bi-RNNLMs and can also be used for lattice rescoring. Experimental results on a meeting transcription task (AMI) show the proposed model consistently outperformed uni-RNNLMs and yield only a slight degradation compared to bi-RNNLMs in N-best rescoring. Additionally, performance improvements can be obtained using lattice rescoring and subsequent confusion network decoding.

Recurrent neural network language models (RNNLMs) have becoming increasingly popular in many applications such as automatic speech recognition (ASR). Significant performance improvements in both perplexity and word error rate over standard n-gram LMs have been widely reported on ASR tasks. In contrast, published research on using RNNLMs for keyword search systems has been relatively limited. In this paper the application of RNNLMs for the IARPA Babel keyword search task is investigated. In order to supplement the limited acoustic transcription data, large amounts of web texts are also used in large vocabulary design and LM training. Various training criteria were then explored to improved RNNLMs' efficiency in both training and evaluation. Significant and consistent improvements on both keyword search and ASR tasks were obtained across all languages.

Recurrent neural network language models (RNNLMs) are powerful language modeling techniques. Significant performance improvements have been reported in a range of tasks including speech recognition compared to n-gram language models. Conventional n-gram and neural network language models are trained to predict the probability of the next word given its preceding context history. In contrast, bidirectional recurrent neural network based language models consider the context from future words as well. This complicates the inference process, but has theoretical benefits for tasks such as speech recognition as additional context information can be used. However to date, very limited or no gains in speech recognition performance have been reported with this form of model. This paper examines the issues of training bidirectional recurrent neural network language models (bi-RNNLMs) for speech recognition. A bi-RNNLM probability smoothing technique is proposed, that addresses the very sharp posteriors that are often observed in these models. The performance of the bi-RNNLMs is evaluated on three speech recognition tasks: broadcast news; meeting transcription (AMI); and low-resource systems (Babel data). On all tasks gains are observed by applying the smoothing technique to the bi-RNNLM. In addition consistent performance gains can be obtained by combining bi-RNNLMs with n-gram and uni-directional RNNLMs.

In recent years, recurrent neural network language models (RNNLMs) have become increasingly popular for a range of applications including speech recognition. However, the training of RNNLMs is computationally expensive, which limits the quantity of data, and size of network, that can be used. In order to fully exploit the power of RNNLMs, efficient training implementations are required. This paper introduces an open-source toolkit, the CUED-RNNLM toolkit, which supports efficient GPU-based training of RNNLMs. RNNLM training with a large number of word level output targets is supported, in contrast to existing tools which used class-based output-targets. Support fotN-best and lattice-based rescoring of both HTK and Kaldi format lattices is included. An example of building and evaluating RNNLMs with this toolkit is presented for a Kaldi based speech recognition system using the AMI corpus. All necessary resources including the source code, documentation and recipe are available online1.

Recently there has been a lot of interest in neural network based language models. These models typically consist of vocabulary dependent input and output layers and one or more vocabulary independent hidden layers. One standard issue with these approaches is that large quantities of training data are needed to ensure robust parameter estimates. This poses a significant problem when only limited data is available. One possible way to address this issue is augmentation: model-based, in the form of language model interpolation, and data-based, in the form of data augmentation. However, these approaches may not always be possible to use due to vocabulary dependent input and output layers. This seriously restricts the nature of the data possible to use in augmentation. This paper describes a general solution whereby only one or more vocabulary independent hidden layers are augmented. Such approach makes it possible to examine augmentation from previously impossible domains. Moreover, this approach paves a direct way for multi-task learning with these models. As a proof of the concept this paper examines the use of multilingual data for augmenting hidden layers of recurrent neural network language models. Experiments are conducted using a set of language packs released within IARPA Babel program.

Recurrent neural network language models (RNNLMs) are becoming increasingly popular for a range of applications including automatic speech recognition. An important issue that limits their possible application areas is the computational cost incurred in training and evaluation. This paper describes a series of new efficiency improving approaches that allows RNNLMs to be more efficiently trained on graphics processing units (GPUs) and evaluated on CPUs. First, a modified RNNLM architecture with a nonclass-based, full output layer structure (F-RNNLM) is proposed. This modified architecture facilitates a novel spliced sentence bunch mode parallelization of F-RNNLM training using large quantities of data on a GPU. Second, two efficient RNNLM training criteria based on variance regularization and noise contrastive estimation are explored to specifically reduce the computation associated with the RNNLM output layer softmax normalisation term. Finally, a pipelined training algorithm utilizing multiple GPUs is also used to further improve the training speed. Initially, RNNLMs were trained on a moderate dataset with 20M words from a large vocabulary conversational telephone speech recognition task. The training time of RNNLM is reduced by up to a factor of 53 on a single GPU over the standard CPU-based RNNLM toolkit. A 56 times speed up in test time evaluation on a CPU was obtained over the baseline F-RNNLMs. Consistent improvements in both recognition accuracy and perplexity were also obtained over C-RNNLMs. Experiments on Google's one billion corpus also reveals that the training of RNNLM scales well.

An important part of the language modelling problem for automatic speech recognition (ASR) systems, and many other related applications, is to appropriately model long-distance context dependencies in natural languages. Hence, statistical language models (LMs) that can model longer span history contexts, for example, recurrent neural network language models (RNNLMs), have become increasingly popular for state-of-the-art ASR systems. As RNNLMs use a vector representation of complete history contexts, they are normally used to rescore N-best lists. Motivated by their intrinsic characteristics, two efficient lattice rescoring methods for RNNLMs are proposed in this paper. The first method uses an n-gram style clustering of history contexts. The second approach directly exploits the distance measure between recurrent hidden history vectors. Both methods produced 1-best performance comparable to a 10 k-best rescoring baseline RNNLM system on two large vocabulary conversational telephone speech recognition tasks for US English and Mandarin Chinese. Consistent lattice size compression and recognition performance improvements after confusion network (CN) decoding were also obtained over the prefix tree structured N-best rescoring approach.

Recurrent neural network language models (RNNLMs) have become an increasingly popular choice for speech and language processing tasks including automatic speech recognition (ASR). As the generalization patterns of RNNLMs and n-gram LMs are inherently different, RNNLMs are usually combined with n-gram LMs via a fixed weighting based linear interpolation in state-of-the-art ASR systems. However, previous work doesn't fully exploit the difference of modelling power of the RNNLMs and n-gram LMs as n-gram level changes. In order to fully exploit the detailed n-gram level complementary attributes between the two LMs, a back-off based compact representation of n-gram dependent interpolation weights is proposed in this paper. This approach allows weight parameters to be robustly estimated on limited data. Experimental results are reported on the three tasks with varying amounts of training data. Small and consistent improvements in both perplexity and WER were obtained using the proposed interpolation approach over the baseline fixed weighting based linear interpolation.

Recurrent neural network language models (RNNLMs) are becoming increasingly popular for speech recognition. Previously, we have shown that RNNLMs with a full (non-classed) output layer (F-RNNLMs) can be trained efficiently using a GPU giving a large reduction in training time over conventional class-based models (C-RNNLMs) on a standard CPU. However, since test-time RNNLM evaluation is often performed entirely on a CPU, standard F-RNNLMs are inefficient since the entire output layer needs to be calculated for normalisation. In this paper, it is demonstrated that C-RNNLMs can be efficiently trained on a GPU, using our spliced sentence bunch technique which allows good CPU test-time performance (42× speedup over F-RNNLM). Furthermore, the performance of different classing approaches is investigated. We also examine the use of variance regularisation of the softmax denominator for F-RNNLMs and show that it allows F-RNNLMs to be efficiently used in test (56× speedup on a CPU). Finally the use of two GPUs for F-RNNLM training using pipelining is described and shown to give a reduction in training time over a single GPU by a factor of 1.6×.

In recent years recurrent neural network language models (RNNLMs) have been successfully applied to a range of tasks including speech recognition. However, an important issue that limits the quantity of data used, and their possible application areas, is the computational cost in training. A signi??cant part of this cost is associated with the softmax function at the output layer, as this requires a normalization term to be explicitly calculated. This impacts both the training and testing speed, especially when a large output vocabulary is used. To address this problem, noise contrastive estimation (NCE) is explored in RNNLM training. NCE does not require the above normalization during both training and testing. It is insensitive to the output layer size. On a large vocabulary conversational telephone speech recognition task, a doubling in training speed on a GPU and a 56 times speed up in test time evaluation on a CPU were obtained.

In this paper we investigate the use of noise-robust features characterizing the speech excitation signal as complementary features to the usually considered vocal tract based features for Automatic Speech Recognition (ASR). The proposed Excitation-based Features (EBF) are tested in a state-of-the-art Deep Neural Network (DNN) based hybrid acoustic model for speech recognition. The suggested excitation features expand the set of periodicity features previously considered for ASR, expecting that these features help in a better discrimination of the broad phonetic classes (e.g., fricatives, nasal, vowels, etc.). Our experiments on the AMI meeting transcription system showed that the proposed EBF yield a relative word error rate reduction of about 5% when combined with conventional PLP features. Further experiments led on Aurora4 confirmed the robustness of the EBF to both additive and convolutive noises, with a relative improvement of 4.3% obtained by combinining them with mel filter banks.

Recurrent neural network language models (RNNLM) have become an increasingly popular choice for state-of-the-art speech recognition systems. Linguistic factors in??uencing the realization of surface word sequences, for example, expressive richness, are only implicitly learned by RNNLMs. Observed sentences and their associated alternative paraphrases representing the same meaning are not explicitly related during training. In order to improve context coverage and generalization, paraphrastic RNNLMs are investigated in this paper. Multiple paraphrase variants were automatically generated and used in paraphrastic RNNLM training. Using a paraphrastic multi-level RNNLM modelling both word and phrase sequences, signi??cant error rate reductions of 0.6% absolute and perplexity reduction of 10% relative were obtained over the baseline RNNLM on a large vocabulary conversational telephone speech recognition system trained on 2000 hours of audio and 545 million words of texts. The overall improvement over the baseline n-gram LM was increased from 8.4% to 11.6% relative.

Recurrent neural network language models (RNNLMs) have recently become increasingly popular for many applications including speech recognition. In previous research RNNLMs have normally been trained on well-matched in-domain data. The adaptation of RNNLMs remains an open research area to be explored. In this paper, genre and topic based RNNLMadaptation techniques are investigated for a multi-genre broadcast transcription task. A number of techniques including Probabilistic Latent Semantic Analysis, Latent Dirichlet Allocation and Hierarchical Dirichlet Processes are used to extract show level topic information. These were then used as additional input to the RNNLM during training, which can facilitate unsupervised test time adaptation. Experiments using a state-of-theart LVCSR system trained on 1000 hours of speech and more than 1 billion words of text showed adaptation could yield perplexity reductions of 8% relatively over the baseline RNNLM and small but consistent word error rate reductions.

Recurrent neural network language models (RNNLM) have become an increasingly popular choice for state-of-the-art speech recognition systems due to their inherently strong generalization performance. As these models use a vector representation of complete history contexts, RNNLMs are normally used to rescore N-best lists. Motivated by their intrinsic characteristics, two novel lattice rescoring methods for RNNLMs are investigated in this paper. The first uses an n-gram style clustering of history contexts. The second approach directly exploits the distance measure between hidden history vectors. Both methods produced 1-best performance comparable with a 10k-best rescoring baseline RNNLM system on a large vocabulary conversational telephone speech recognition task. Significant lattice size compression of over 70% and consistent improvements after confusion network (CN) decoding were also obtained over the N-best rescoring approach.

Over the past few decades, a range of front-end techniques have been proposed to improve the robustness of automatic speech recognition systems against environmental distortion. While these techniques are effective for small tasks consisting of carefully designed data sets, especially when used with a classical acoustic model, there has been limited evidence that they are useful for a state-of-the-art system with large scale realistic data. This paper focuses on reverberation as a type of distortion and investigates the degree to which dereverberation processing can improve the performance of various forms of acoustic models based on deep neural networks (DNNs) in a challenging meeting transcription task using a single distant microphone. Experimental results show that dereverberation improves the recognition performance regardless of the acoustic model structure and the type of the feature vectors input into the neural networks, providing additional relative improvements of 4.7% and 4.1% to our best configured speaker-independent and speaker-adaptive DNN-based systems, respectively.

Meeting transcription is a very useful and challenging task. The majority of research to date has focused on individual meeting, or only a small group of meetings. In many practical deploy-ments, multiple related meetings will take place over a long pe-riod of time. This paper describes an initial investigation of how this long-term data can be used to improve meeting tran-scription. A corpus of technical meetings, using a single micro-phone array, was collected over a two year period, yielding a total of 179 hours of meeting data. Baseline systems using deep neural network acoustic models, in both Tandem and Hybrid configurations, and neural network-based language models are described. The impact of supervised and unsupervised adap-tation of the acoustic models is then evaluated, as well as the impact of improved language models.

Recurrent neural network language models (RNNLMs) are be-coming increasingly popular for a range of applications includ-ing speech recognition. However, an important issue that limits the quantity of data, and hence their possible application ar-eas, is the computational cost in training. A standard approach to handle this problem is to use class-based outputs, allowing systems to be trained on CPUs. This paper describes an alter-native approach that allows RNNLMs to be efficiently trained on GPUs. This enables larger quantities of data to be used, and networks with an unclustered, full output layer to be trained. To improve efficiency on GPUs, multiple sentences are “spliced” together for each mini-batch or “bunch ” in training. On a large vocabulary conversational telephone speech recognition task, the training time was reduced by a factor of 27 over the stan-dard CPU-based RNNLM toolkit. The use of an unclustered, full output layer also improves perplexity and recognition per-formance over class-based RNNLMs. Index Terms: language models, recurrent neural network, speech recognition, GPU

The Context-Dependent Deep-Neural-Network HMM, or CD-DNN-HMM, is a recently proposed acoustic-modeling tech-nique for HMM-based speech recognition that can greatly out-perform conventional Gaussian-mixture based HMMs. For ex-ample, a CD-DNN-HMM trained on the 2000h Fisher corpus achieves 14.4 % word error rate on the Hub5’00-FSH speaker-independent phone-call transcription task, compared to 19.6% obtained by a state-of-the-art, conventional discriminatively trained GMM-based HMM. That CD-DNN-HMM, however, took 59 days to train on a modern GPGPU—the immense computational cost of the mini-batch based back-propagation (BP) training is a major road-block. Unlike the familiar Baum-Welch training for conven-tional HMMs, BP cannot be efficiently parallelized across data. In this paper we show that the pipelined approximation to BP, which parallelizes computation with respect to layers, is an efficient way of utilizing multiple GPGPU cards in a single server. Using 2 and 4 GPGPUs, we achieve a 1.9 and 3.3 times end-to-end speed-up, at parallelization efficiency of 0.95 and 0.82, respectively, at no loss of recognition accuracy. Index Terms: speech recognition, deep neural networks, paral-lelization, GPGPU

We investigate the potential of Context-Dependent Deep-Neural-Network HMMs, or CD-DNN-HMMs, from a feature-engineering perspective. Recently, we had shown that for speaker-independent transcription of phone calls (NIST RT03S Fisher data), CD-DNN-HMMs reduced the word error rate by as much as one third-from 27.4%, obtained by discriminatively trained Gaussian-mixture HMMs with HLDA features, to 18.5%-using 300+ hours of training data (Switchboard), 9000+ tied triphone states, and up to 9 hidden network layers.