Although Speech Large Language Models have achieved notable progress, a substantial modality reasoning gap remains: their reasoning performance on speech inputs is markedly weaker than on text.This gap could be associated with representational drift across Transformer layers and behavior deviations in long-chain reasoning.To address this issue, we introduce TARS, a reinforcement-learning framework that aligns text-conditioned and speech-conditioned trajectories through an asymmetric reward design.The framework employs two dense and complementary signals: representation alignment, which measures layer-wise hidden-state similarity between speech-and text-conditioned trajectories, and behavior alignment, which evaluates semantic consistency between generated outputs and reference text completions.Experiments on challenging reasoning benchmarks, including MMSU and OBQA, show that our approach significantly narrows the modality reasoning gap and achieves state-of-the-art performance among 7B-scale Speech LLMs.

Hui Wang, Jinghua Zhao, Yifan Yang, Shujie Liu, Junyang Chen, Yanzhe Zhang, Shiwan Zhao, Jinyu Li, Jiaming Zhou, Haoqin Sun, Yan Lu, Yong Qin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Lingwei Meng, Long Zhou, Shujie Liu, Sanyuan Chen, Bing Han, Shujie Hu, Yanqing Liu, Jinyu Li, Sheng Zhao, Xixin Wu, Helen M. Meng, Furu Wei. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Speech-aware language models (LMs) have demonstrated capabilities in understanding spoken language while generating text-based responses. However, enabling them to produce speech output efficiently and effectively remains a challenge. In this paper, we present SLM-S2ST, a multimodal LM for direct speech-to-speech translation (S2ST), built on the opensource Phi4-MM model. SLM-S2ST extends its predecessor by generating translated speech using an audio transformer head that predicts audio tokens with a delay relative to text tokens, followed by a streaming vocoder for waveform synthesis. Our experimental results on the CVSS-C dataset demonstrate SLMS2ST’s superior performance, significantly surpassing existing baseline models trained on the same dataset. Furthermore, when we scale up the training data and the model size, SLM-S2ST reaches on-par performance with the current SOTA model.

End-to-End Automatic Speech Recognition (ASR) has advanced significantly yet still struggles with rare and domain-specific entities. This paper introduces a simple yet efficient prompt-based biasing technique for contextualized ASR, enhancing recognition accuracy by leverage a unified multi-task learning framework. The approach comprises two key components: a prompt biasing model which is trained to determine when to focus on entities in prompt, and a entity filtering mechanism which efficiently filters out irrelevant entities. Our method significantly enhances ASR accuracy on entities, achieving a relative $30.7 \%$ and $18.0 \%$ reduction in Entity Word Error Rate compared to the baseline model with shallow fusion on in-house domain dataset with small and large entity lists, respectively. The primary advantage of this method lies in its efficiency and simplicity without any structure change, making it lightweight and highly efficient.

Speech language models (Speech LMs) enable end-to-end speech-text modeling within a single model, offering a promising direction for spoken dialogue systems. The choice of speech-text jointly decoding paradigm plays a critical role in performance, efficiency, and alignment quality. In this work, we systematically compare representative joint speech-text decoding strategies—including the interleaved, and parallel generation paradigms—under a controlled experimental setup using the same base language model, speech tokenizer and training data. Our results show that the interleaved approach achieves the best alignment. However it suffers from slow inference due to long token sequence length. To address this, we propose a novel early-stop interleaved (ESI) pattern that not only significantly accelerates decoding but also yields slightly better performance. Additionally, we curate high-quality question answering (QA) datasets to further improve speech QA performance.

In this paper, we introduce V2SFlow, a novel Video-to-Speech (V2S) framework designed to generate natural and intelligible speech directly from silent talking face videos. While recent V2S systems have shown promising results on constrained datasets with limited speakers and vocabularies, their performance often degrades on real-world, unconstrained datasets due to the inherent variability and complexity of speech signals. To address these challenges, we decompose the speech signal into manageable subspaces (content, pitch, and speaker information), each representing distinct speech attributes, and predict them directly from the visual input. To generate coherent and realistic speech from these predicted attributes, we employ a rectified flow matching decoder built on a Transformer architecture, which models efficient probabilistic pathways from random noise to the target speech distribution. Extensive experiments demonstrate that V2SFlow significantly outperforms state-of-the-art methods, even surpassing the naturalness of ground truth utterances.

Large language models (LLMs) have made significant advancements in natural language processing and are concurrently extending the language ability to other modalities, such as speech and vision. Nevertheless, most of the previous work focuses on prompting LLMs with perception abilities like auditory comprehension, and the effective approach for augmenting LLMs with speech synthesis capabilities remains ambiguous. In this paper, we conduct a comprehensive empirical exploration of boosting LLMs with the ability to generate speech, by combining pre-trained LLM LLaMA/OPT and text-to-speech synthesis model VALL-E. We compare three integration methods between LLMs and speech synthesis models, including directly fine-tuned LLMs, superposed layers of LLMs and VALL-E, and coupled LLMs and VALL-E using LLMs as a powerful text encoder. Experimental results show that, using LoRA method to fine-tune LLMs directly to boost the speech synthesis capability does not work well, and superposed LLMs and VALL-E can improve the quality of generated speech both in speaker similarity and word error rate (WER). Among these three methods, coupled methods leveraging LLMs as the text encoder can achieve the best performance, making it outperform original speech synthesis models with a consistently better speaker similarity and a significant (10.9%) WER reduction.

Obtaining word timestamp information from end-to-end (E2E) ASR models remains challenging due to the lack of explicit time alignment during training. This issue is further complicated in multilingual models. Existing methods, either rely on lexicons or introduce additional tokens, leading to scalability issues and increased computational costs. In this work, we propose a new approach to estimate word boundaries without relying on lexicons. Our method leverages word embeddings from sub-word token units and a pretrained ASR model, requiring only word alignment information during training. Our proposed method can scale-up to any number of languages without incurring any additional cost. We validate our approach using a multilingual ASR model trained on five languages and demonstrate its effectiveness against a strong baseline.

Text-to-video (T2V) models have recently undergone rapid and substantial advancements. Nevertheless, due to limitations in data and computational resources, achieving efficient generation of long videos with rich motion dynamics remains a significant challenge. To generate high-quality, dynamic, and temporally consistent long videos, this paper presents ARLON, a novel framework that boosts diffusion Transformers with autoregressive (\textbf{AR}) models for long (\textbf{LON}) video generation, by integrating the coarse spatial and long-range temporal information provided by the AR model to guide the DiT model effectively. Specifically, ARLON incorporates several key innovations: 1) A latent Vector Quantized Variational Autoencoder (VQ-VAE) compresses the input latent space of the DiT model into compact and highly quantized visual tokens, bridging the AR and DiT models and balancing the learning complexity and information density; 2) An adaptive norm-based semantic injection module integrates the coarse discrete visual units from the AR model into the DiT model, ensuring effective guidance during video generation; 3) To enhance the tolerance capability of noise introduced from the AR inference, the DiT model is trained with coarser visual latent tokens incorporated with an uncertainty sampling module. Experimental results demonstrate that ARLON significantly outperforms the baseline OpenSora-V1.2 on eight out of eleven metrics selected from VBench, with notable improvements in dynamic degree and aesthetic quality, while delivering competitive results on the remaining three and simultaneously accelerating the generation process. In addition, ARLON achieves state-of-the-art performance in long video generation, outperforming other open-source models in this domain. Detailed analyses of the improvements in inference efficiency are presented, alongside a practical application that demonstrates the generation of long videos using progressive text prompts. Project page: \url{http://aka.ms/arlon}.

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.

To advance continuous token modeling and temporal-coherence enforcement, we propose FELLE, an autoregressive model that integrates language modeling with token-wise flow matching. By leveraging the autoregressive nature of language models and the generative efficacy of flow matching, FELLE effectively predicts continuous-valued tokens (mel-spectrograms). For each continuous-valued token, FELLE modifies the general prior distribution in flow matching by incorporating information from the previous step, improving coherence and stability. Furthermore, to enhance synthesis quality, FELLE introduces a coarse-to-fine flow-matching mechanism, generating continuous-valued tokens hierarchically, conditioned on the language model's output. Experimental results demonstrate the potential of incorporating flow-matching techniques in autoregressive mel-spectrogram modeling, leading to significant improvements in TTS generation quality, as shown in https://aka.ms/felle.

Generating natural-sounding, multi-speaker dialogue is crucial for applications such as podcast creation, virtual agents, and multimedia content generation. However, existing systems struggle to maintain speaker consistency, model overlapping speech, and synthesize coherent conversations efficiently. In this paper, we introduce CoVoMix2, a fully non-autoregressive framework for zero-shot multi-talker dialogue generation. CoVoMix2 directly predicts mel-spectrograms from multi-stream transcriptions using a flow-matching-based generative model, eliminating the reliance on intermediate token representations. To better capture realistic conversational dynamics, we propose transcription-level speaker disentanglement, sentence-level alignment, and prompt-level random masking strategies. Our approach achieves state-of-the-art performance, outperforming strong baselines like MoonCast and Sesame in speech quality, speaker consistency, and inference speed. Notably, CoVoMix2 operates without requiring transcriptions for the prompt and supports controllable dialogue generation, including overlapping speech and precise timing control, demonstrating strong generalizability to real-world speech generation scenarios. Audio samples are available at https://www.microsoft.com/en-us/research/project/covomix/covomix2.

Recent advances in zero-shot text-to-speech (TTS) synthesis have achieved high-quality speech generation for unseen speakers, but most systems remain unsuitable for real-time applications because of their offline design. Current streaming TTS paradigms often rely on multi-stage pipelines and discrete representations, leading to increased computational cost and suboptimal system performance. In this work, we propose StreamMel, a pioneering single-stage streaming TTS framework that models continuous mel-spectrograms. By interleaving text tokens with acoustic frames, StreamMel enables low-latency, autoregressive synthesis while preserving high speaker similarity and naturalness. Experiments on LibriSpeech demonstrate that StreamMel outperforms existing streaming TTS baselines in both quality and latency. It even achieves performance comparable to offline systems while supporting efficient real-time generation, showcasing broad prospects for integration with real-time speech large language models. Audio samples are available at:https://aka.ms/StreamMel.

The growing need for instant spoken language transcription and translation is driven by increased global communication and cross-lingual interactions. This has made offering translations in multiple languages essential for user applications. Traditional approaches to automatic speech recognition (ASR) and speech translation (ST) have often relied on separate systems, leading to inefficiencies in computational resources, and increased synchronization complexity in real time. In this paper, we propose a streaming Transformer-Transducer (T-T) model able to jointly produce many-to-one and one-to-many transcription and translation using a single decoder. We introduce a novel method for joint token-level serialized output training based on timestamp information to effectively produce ASR and ST outputs in the streaming setting. Experiments on {it,es,de}↔en prove the effectiveness of our approach, enabling the generation of one-to-many joint outputs with a single decoder for the first time.

Memory constraint of always-on devices is one of the major concerns when deploying speech processing models on these devices. While larger models trained with sufficiently large amount of data generally perform better, making them fit in the device memory is a demanding challenge. In this paper, we aim to reduce model size by reparameterizing model weights across Transformer encoder layers and assuming a special weight composition and structure. More specifically, inspired by ResNet [1] and the more recent LoRA [2] work, we propose an approach named ResidualTransformer, where each weight matrix in a Transformer layer comprises 1) a shared full-rank component with its adjacent layers, and 2) a unique low-rank component to itself. The low-rank matrices only account for a small amount of model size increase. In addition, we add diagonal weight matrices to improve modeling capacity of the low-rank matrices. Experiments of our 10k-hour speech recognition and speech translation tasks show that the Transformer encoder size can be reduced by ∼3× with very slight performance degradation.

Token-level serialized output training (t-SOT) was recently proposed to address the challenge of streaming multi-talker automatic speech recognition (ASR). T-SOT effectively handles overlapped speech by representing multi-talker transcriptions as a single token stream with ⟨cc⟩ symbols interspersed. However, the use of a naive neural transducer architecture significantly constrained its applicability for text-only adaptation. To overcome this limitation, we propose a novel t-SOT model structure that incorporates the idea of factorized neural transducers (FNT). The proposed method separates a language model (LM) from the transducer’s predictor and handles the unnatural token order resulting from the use of ⟨cc⟩ symbols in t-SOT. We achieve this by maintaining multiple hidden states and introducing special handling of the ⟨cc⟩ tokens within the LM. The proposed t-SOT FNT model achieves comparable performance to the original t-SOT model while retaining the ability to reduce word error rate (WER) on both single and multi-talker datasets through text-only adaptation.

End-to-end speech translation (ST) for conversation recordings involves several under-explored challenges such as speaker diarization (SD) without accurate word time stamps and handling of overlapping speech in a streaming fashion. In this work, we propose DiariST, the first streaming ST and SD solution. It is built upon a neural transducer-based streaming ST system and integrates tokenlevel serialized output training and t-vector, which were originally developed for multi-talker speech recognition. Due to the absence of evaluation benchmarks in this area, we develop a new evaluation dataset, DiariST-AliMeeting, by translating the reference Chinese transcriptions of the AliMeeting corpus into English. We also propose new metrics, called speaker-agnostic BLEU and speaker-attributed BLEU, to measure the ST quality while taking SD accuracy into account. Our system achieves a strong ST and SD capability compared to offline systems based on Whisper, while performing streaming inference for overlapping speech. To facilitate the research in this new direction, we release the evaluation data, the offline baseline systems, and the evaluation code.

While recent large-scale text-to-speech (TTS) models have achieved significant progress, they still fall shorts in speech quality, similarity, and prosody. Considering that speech intricately encompasses various attributes (e.g., content, prosody, timbre, and acoustic details) that pose significant challenges for generation, a natural idea is to factorize speech into individual subspaces representing different attributes and generate them individually. Motivated by it, we propose a TTS system with novel factorized diffusion models to generate natural speech in a zero-shot way. Specifically, 1) we design a neural codec with factorized vector quantization (FVQ) to disentangle speech waveform into subspaces of content, prosody, timbre, and acoustic details; 2) we propose a factorized diffusion model, which generates attributes in each subspace following its corresponding prompt. With this factorization design, our method can effectively and efficiently model the intricate speech with disentangled subspaces in a divide-and-conquer way. Experimental results show that our method outperforms the state-of-the-art TTS systems on quality, similarity, prosody, and intelligibility.

Accurate control of the total duration of generated speech by adjusting the speech rate is crucial for various text-to-speech (TTS) applications.However, the impact of adjusting the speech rate on speech quality, such as intelligibility and speaker characteristics, has been underexplored.In this work, we propose a novel total-duration-aware (TDA) duration model for TTS, where phoneme durations are predicted not only from the text input but also from an additional input of the total target duration.We also propose a MaskGIT-based duration model that enhances the diversity and quality of the predicted phoneme durations.Our results demonstrate that the proposed TDA duration models achieve better intelligibility and speaker similarity for various speech rate configurations compared to the baseline models.We also show that the proposed MaskGIT-based model can generate phoneme durations with higher quality and diversity compared to its regression or flow-matching counterparts.

There is a rising interest and trend in research towards directly translating speech from one language to another, known as end-to-end speech-to-speech translation. However, most end-to-end models struggle to outperform cascade models, i.e., a pipeline framework by concatenating speech recognition, machine translation and text-to-speech models. The primary challenges stem from the inherent complexities involved in direct translation tasks and the scarcity of data. In this study, we introduce a novel model framework TransVIP that leverages diverse datasets in a cascade fashion yet facilitates end-to-end inference through joint probability. Furthermore, we propose two separated encoders to preserve the speaker’s voice characteristics and isochrony from the source speech during the translation process, making it highly suitable for scenarios such as video dubbing. Our experiments on the French-English language pair demonstrate that our model outperforms the current state-of-the-art speech-to-speech translation model.

Recent advancements in zero-shot text-to-speech (TTS) modeling have led to significant strides in generating high-fidelity and diverse speech. However, dialogue generation, along with achieving human-like naturalness in speech, continues to be a challenge. In this paper, we introduce CoVoMix: Conversational Voice Mixture Generation, a novel model for zero-shot, human-like, multi-speaker, multi-round dialogue speech generation. CoVoMix first converts dialogue text into multiple streams of discrete tokens, with each token stream representing semantic information for individual talkers. These token streams are then fed into a flow-matching based acoustic model to generate mixed mel-spectrograms. Finally, the speech waveforms are produced using a HiFi-GAN model. Furthermore, we devise a comprehensive set of metrics for measuring the effectiveness of dialogue modeling and generation. Our experimental results show that CoVoMix can generate dialogues that are not only human-like in their naturalness and coherence but also involve multiple talkers engaging in multiple rounds of conversation. This is exemplified by instances generated in a single channel where one speaker's utterance is seamlessly mixed with another's interjections or laughter, indicating the latter's role as an attentive listener. Audio samples are enclosed in the supplementary.

Neural audio codec tokens serve as the fundamental building blocks for speech language model (SLM)-based speech generation. However, there is no systematic understanding on how the codec system affects the speech generation performance of the SLM. In this work, we examine codec tokens within SLM framework for speech generation to provide insights for effective codec design. We retrain existing high-performing neural codec models on the same data set and loss functions to compare their performance in a uniform setting. We integrate codec tokens into two SLM systems: masked-based parallel speech generation system and an auto-regressive (AR) plus non-auto-regressive (NAR) model-based system. Our findings indicate that better speech reconstruction in codec systems does not guarantee improved speech generation in SLM. A high-quality codec decoder is crucial for natural speech production in SLM, while speech intelligibility depends more on quantization mechanism.

People change their tones of voice, often accompanied by nonverbal vocalizations (NVs) such as laughter and cries, to convey rich emotions. However, most text-to-speech (TTS) systems lack the capability to generate speech with rich emotions, including NVs. This paper introduces EmoCtrl-TTS, an emotion-controllable zeroshot TTS that can generate highly emotional speech with NVs for any speaker. EmoCtrl-TTS leverages arousal and valence values, as well as laughter embeddings, to condition the flow-matchingbased zero-shot TTS. To achieve high-quality emotional speech generation, EmoCtrl-TTS is trained using more than 27,000 hours of expressive data curated based on pseudo-labeling. Comprehensive evaluations demonstrate that EmoCtrl-TTS excels in mimicking the emotions of audio prompts in speech-to-speech translation scenarios. We also show that EmoCtrl-TTS can capture emotion changes, express strong emotions, and generate various NVs in zero-shot TTS. See https://aka.ms/emoctrl-tts for demo samples.

Models for streaming speech translation (ST) can achieve high accuracy and low latency if they’re developed with vast amounts of paired audio in the source language and written text in the target language. Yet, these text labels for the target language are often pseudo labels due to the prohibitive cost of manual ST data labeling. In this paper, we introduce a methodology named Connectionist Temporal Classification guided modality matching (CTC-GMM) that enhances the streaming ST model by leveraging extensive machine translation (MT) text data. This technique employs CTC to compress the speech sequence into a compact embedding sequence that matches the corresponding text sequence, allowing us to utilize matched source-target language text pairs from the MT corpora to refine the streaming ST model further. Our evaluations with FLEURS and CoVoST2 show that the CTC-GMM approach can increase translation accuracy relatively by 13.9% and 6.4% respectively, while also boosting decoding speed by 59.7% on GPU.

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.

Recent advancements in generative speech models based on audio-text prompts have enabled remarkable innovations like high-quality zero-shot text-to-speech. However, existing models still face limitations in handling diverse audio-text speech generation tasks involving transforming input speech and processing audio captured in adverse acoustic conditions. This paper introduces SpeechX, a versatile speech generation model capable of zero-shot TTS and various speech transformation tasks, dealing with both clean and noisy signals. SpeechX combines neural codec language modeling with multi-task learning using task-dependent prompting, enabling unified and extensible modeling and providing a consistent way for leveraging textual input in speech enhancement and transformation tasks. Experimental results show SpeechX's efficacy in various tasks, including zero-shot TTS, noise suppression, target speaker extraction, speech removal, and speech editing with or without background noise, achieving comparable or superior performance to specialized models across tasks.

Recent research shows a big convergence in model architecture, training objectives, and inference methods across various tasks for different modalities. In this paper, we proposeVioLA, a single auto-regressive Transformer decoder-only network that unifies various cross-modal tasks involving speech and text, such as speech-to-text, text-to-text, text-to-speech, and speech-to-speech tasks, as a conditional language model task via multi-task learning framework. To accomplish this, we first convert the speech utterances to discrete tokens (similar to the textual data) using an offline neural codec encoder. In such a way, all these tasks are converted to token-based sequence prediction problems, which can be naturally handled with one conditional language model. We further integrate task IDs (TID), language IDs (LID), and LSTM-based acoustic embedding into the proposed model to enhance the modeling capability of handling different languages and tasks. Experimental results demonstrate that the proposedVioLAmodel can support both single-modal and cross-modal tasks well, and the decoder-only model achieves a comparable and even better performance than the strong baselines.

How to boost speech pre-training with textual data is an unsolved problem due to the fact that speech and text are very different modalities with distinct characteristics. In this paper, we propose a cross-modalSpeechandLanguageModel (SpeechLM) to explicitly align speech and text pre-training with a pre-defined unified discrete representation. Specifically, we introduce two alternative discrete tokenizers to bridge the speech and text modalities, including phoneme-unit and hidden-unit tokenizers, which can be trained using unpaired speech or a small amount of paired speech-text data. Based on the trained tokenizers, we convert the unlabeled speech and text data into tokens of phoneme units or hidden units. The pre-training objective is designed to unify the speech and the text into the same discrete semantic space with a unified Transformer network. We evaluate SpeechLM on various spoken language processing tasks including speech recognition, speech translation, and universal representation evaluation framework SUPERB, demonstrating significant improvements on content-related tasks. Code and models are available athttps://aka.ms/SpeechLM.

Although speech is a simple and effective way for humans to communicate with the outside world, a more realistic speech interaction contains multimodal information, e.g., vision, text. How to design a unified framework to integrate different modal information and leverage different resources (e.g., visual-audio pairs, audio-text pairs, unlabeled speech, and unlabeled text) to facilitate speech representation learning was not well explored. In this paper, we propose a unified cross-modal representation learning frameworkVatLM(Visual-Audio-Text Language Model). The proposedVatLMemploys a unified backbone network to model the modality-independent information and utilizes three simple modality-dependent modules to preprocess visual, speech, and text inputs. In order to integrate these three modalities into one shared semantic space,VatLMis optimized with a masked prediction task of unified tokens, given by our proposed unified tokenizer. We evaluate the pre-trainedVatLMon audio-visual related downstream tasks, including audio-visual speech recognition (AVSR), and visual speech recognition (VSR) tasks. Results show that the proposedVatLMoutperforms previous state-of-the-art models, such as the audio-visual pre-trained AV-HuBERT model, and analysis also demonstrates thatVatLMis capable of aligning different modalities into the same space. To facilitate future research, we release the code and pre-trained models athttps://aka.ms/vatlm.

Simultaneous Speech-to-Text translation serves a critical role in real-time crosslingual communication. Despite the advancements in recent years, challenges remain in achieving stability in the translation process, a concern primarily manifested in the flickering of partial results. In this paper, we propose a novel revision-controllable method designed to address this issue. Our method introduces an allowed revision window within the beam search pruning process to screen out candidate translations likely to cause extensive revisions, leading to a substantial reduction in flickering and, crucially, providing the capability to completely eliminate flickering. The experiments demonstrate the proposed method can significantly improve the decoding stability without compromising substantially on the translation quality.

The integration of Language Models (LMs) has proven to be an effective way to address domain shifts in speech recognition. However, these approaches usually require a significant amount of target domain text data for the training of LMs. Different from these methods, in this work, with only a domain-specific text prompt, we propose two zero-shot ASR domain adaptation methods using LLaMA, a 7-billionparameter large language model (LLM). LLM is used in two ways: 1) second-pass rescoring: reranking N-best hypotheses of a given ASR system with LLaMA; 2) deep LLM-fusion: incorporating LLM into the decoder of an encoder-decoder based ASR system. Experiments show that, with only one domain prompt, both methods can effectively reduce word error rates (WER) on out-of-domain TedLium-2 and SPGISpeech datasets. Especially, the deep LLM-fusion has the advantage of better recall of entity and out-of-vocabulary words.

In real-world applications, users often require both translations and transcriptions of speech to enhance their comprehension, particularly in streaming scenarios where incremental generation is necessary. This paper introduces a streaming Transformer-Transducer that jointly generates automatic speech recognition (ASR) and speech translation (ST) outputs using a single decoder. To produce ASR and ST content effectively with minimal latency, we propose a joint token-level serialized output training method that interleaves source and target words by leveraging an off-the-shelf textual aligner. Experiments in monolingual (it-en) and multilingual ({de,es,it}-en) settings demonstrate that our approach achieves the best quality-latency balance. With an average ASR latency of 1s and ST latency of $1.3 \mathrm{~s}$, our model shows no degradation or even improves output quality compared to separate ASR and ST models, yielding an average improvement of 1.1 WER and 0.4 BLEU in the multilingual case.

We propose gated language experts and curriculum training to enhance multilingual transformer transducer models without requiring user input for language identification (LID) during inference. Our method incorporates a gating mechanism and LID loss to enable transformer experts to learn language-specific information. Linear experts are applied on joint network to stabilize training. The curriculum training scheme leverages LID to guide gated experts in improving their respective language-specific performance. Experimental results on an English and Spanish bilingual task show significant average relative word error reductions of 12.5 % and 7.3 % compared to the baseline bilingual and monolingual models, respectively. Our models even perform similarly to upper-bound models with oracle LID. Extending our approach to trilingual, quadrilingual, and pentalingual models reveals similar advantages to those seen in the bilingual models, highlighting its ease of extension to multiple languages.

Large language models (LLMs) have achieved remarkable success in the field of natural language processing, enabling better human-computer interaction using natural language. However, the seamless integration of speech signals into LLMs has not been explored well. The “decoder-only“ architecture has also not been well studied for speech processing tasks. In this research, we introduce Speech-LLaMA, a novel approach that effectively incorporates acoustic information into text-based large language models. Our method leverages Connectionist Temporal Classification and a simple audio encoder to map the compressed acoustic features to the continuous semantic space of the LLM. In addition, we further probe the decoder-only architecture for speech-to-text tasks by training a smaller scale randomly initialized speech-LLaMA model from speech-text paired data alone. We conduct experiments on multilingual speech-to-text translation tasks and demonstrate a significant improvement over strong baselines, highlighting the potential advantages of decoder-only models for speech-to-text conversion.

Streaming automatic speech recognition (ASR) and speech translation (ST) tasks have extensively utilized neural transducers. In this paper, we present our endeavor to construct a Streaming Multilingual Speech Model (SM2), which employs a single neural transducer model for transcribing or translating multiple languages into target languages. SM2is trained using weakly supervised data created by converting speech recognition transcriptions with a machine translation model. Leveraging 351 thousand hours of speech training data from 25 languages, SM2achieves impressive ST performance. Furthermore, we demonstrate the truly zero-shot capability of SM2when expanding to new target languages, generating high-quality zero-shot ST translation for {source-speech, target-text} pairs that were not seen during training.

Self-supervised learning (SSL) methods such as WavLM have shown promising speech separation (SS) results in small-scale simulation-based experiments. In this work, we extend the exploration of the SSL-based SS by massively scaling up both the pre-training data (more than 300K hours) and fine-tuning data (10K hours). We also investigate various techniques to efficiently integrate the pre-trained model with the SS network under a limited computation budget, including a low frame rate SSL model training setup and a fine-tuning scheme using only the part of the pre-trained model. Compared with a supervised baseline and the WavLM-based SS model using feature embeddings obtained with the previously released 94K hours trained WavLM, our proposed model obtains 15.9% and 11.2% of relative word error rate (WER) reductions, respectively, for a simulated far-field speech mixture test set. For conversation transcription on real meeting recordings using continuous speech separation, the proposed model achieves 6.8% and 10.6% of relative WER reductions over the purely supervised baseline on AMI and ICSI evaluation sets, respectively, while reducing the computational cost by 38%.

In this work, we present a simple but effective method, CTCBERT, for advancing hidden-unit BERT (HuBERT). HuBERT applies a frame-level cross-entropy (CE) loss, which is similar to most acoustic model training. However, CTCBERT performs the model training with the Connectionist Temporal Classification (CTC) objective after removing duplicated IDs in each masked region. The idea stems from the observation that there can be significant errors in alignments when using clustered or aligned IDs. CTC learns alignments implicitly, indicating that learning with CTC can be more flexible when misalignment exists. We examine CTCBERT on IDs from HuBERT Iter1, HuBERT Iter2, and PBERT. The CTC training brings consistent improvements compared to the CE training. Furthermore, when loading blank-related parameters during finetuning, slight improvements are observed. Evaluated on the Librispeech 960-100h setting, the relative WER improvements of CTCBERT are 2%-11% over HuBERT and PERT on test-other data.

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.

Self-supervised learning (SSL), which utilizes the input data itself for representation learning, has achieved state-of-the-art results for various downstream speech tasks. However, most of the previous studies focused on offline single-talker applications, with limited investigations in multi-talker cases, especially for streaming scenarios. In this paper, we investigate SSL for streaming multi-talker speech recognition, which generates transcriptions of overlapping speakers in a streaming fashion. Firstly, we observe that conventional SSL techniques do not work well on this task due to the poor representation of overlapping speech. We then propose a novel SSL training objective, referred to as bi-label masked speech prediction, which explicitly preserves representations of all speakers in overlapping speech. We investigate various aspects of the proposed system, including data configuration and quantizer selection. The proposed SSL setup achieves substantially better word error rates on the LibriSpeechMix dataset.

This paper presents a novel streaming automatic speech recognition (ASR) framework for multi-talker overlapping speech captured by a distant microphone array with an arbitrary geometry. Our framework, named t-SOT-VA, capitalizes on independently developed two recent technologies; array-geometry-agnostic continuous speech separation, or VarArray, and streaming multi-talker ASR based on token-level serialized output training (t-SOT). To combine the best of both technologies, we newly design a t-SOT-based ASR model that generates a serialized multi-talker transcription based on two separated speech signals from VarArray. We also propose a pre-training scheme for such an ASR model where we simulate VarArray’s output signals based on monaural single-talker ASR training data. Conversation transcription experiments using the AMI meeting corpus show that the system based on the proposed framework significantly outperforms conventional ones. Our system achieves the state-of-the-art word error rates of 13.7% and 15.5% for the AMI development and evaluation sets, respectively, in the multiple-distant-microphone setting while retaining the streaming inference capability.

We previously proposed contextual spelling correction (CSC) to correct the output of end-to-end (E2E) automatic speech recognition (ASR) models with contextual information such as name, place, etc. Although CSC has achieved reasonable improvement in the biasing problem, there are still two drawbacks for further accuracy improvement. First, due to information limitation in text only hypothesis or weak performance of ASR model on rare domains, the CSC model may fail to correct phrases with similar pronunciation or anti-context cases where all biasing phrases are not present in the utterance. Second, there is a discrepancy between the training and inference of CSC. The bias list in training is randomly selected but in inference there may be more similarity between ground truth phrase and other phrases. To solve above limitations, in this paper we propose an improved non-autoregressive (NAR) spelling correction model for contextual biasing in E2E neural transducer-based ASR systems to improve the previous CSC model from two perspectives: Firstly, we incorporate acoustics information with an external attention as well as text hypotheses into CSC to better distinguish target phrase from dissimilar or irrelevant phrases. Secondly, we design a semantic aware data augmentation schema in training phrase to reduce the mismatch between training and inference to further boost the biasing accuracy. Experiments show that the improved method outperforms the baseline ASR+Biasing system by as much as 20.3% relative name recall gain and achieves stable improvement compared to the previous CSC method over different bias list name coverage ratio.

Direct speech-to-speech translation (S2ST) is an attractive research topic with many advantages compared to cascaded S2ST. However, direct S2ST suffers from the data scarcity problem because the corpora from the speech of the source language to the speech of the target language are very rare. To address this issue, we propose in this paper a Speech2S model, which is jointly pre-trained with unpaired speech and bilingual text data for direct speech-to-speech translation tasks. By effectively leveraging the paired text data, Speech2S is capable of modeling the cross-lingual speech conversion from source to target language. We verify the performance of the proposed Speech2S on Europarl-ST and VoxPopuli datasets. Experimental results demonstrate that Speech2S gets an improvement of about 5 BLEU scores compared to encoder-only pre-training models, and achieves a competitive or even better performance than existing state-of-the-art models1.

Speaker change detection (SCD) is an important feature that improves the readability of the recognized words from an automatic speech recognition (ASR) system by breaking the word sequence into paragraphs at speaker change points. Existing SCD solutions either require additional ensemble for the time based decisions and recognized word sequences, or implement a tight integration between ASR and SCD, limiting the potential optimum performance for both tasks. To address these issues, we propose a novel framework for the SCD task, where an additional SCD module is built on top of an existing Transformer Transducer ASR (TT-ASR) network. Two variants of the SCD network are explored in this framework that naturally estimate speaker change probability for each word, while allowing the ASR and SCD to have independent optimization scheme for the best performance. Experiments show that our methods can significantly improve the F1 score on LibriCSS and Microsoft call center data sets without ASR degradation, compared with a joint SCD and ASR baseline.

Multi-talker automatic speech recognition (ASR) has been studied to generate transcriptions of natural conversation including over-lapping speech of multiple speakers. Due to the difficulty in acquiring real conversation data with high-quality human transcriptions, a naïve simulation of multi-talker speech by randomly mixing multiple utterances was conventionally used for model training. In this work, we propose an improved technique to simulate multi-talker overlap-ping speech with realistic speech overlaps, where an arbitrary pattern of speech overlaps is represented by a sequence of discrete tokens. With this representation, speech overlapping patterns can be learned from real conversations based on a statistical language model, such as N-gram, which can be then used to generate multi-talker speech for training. In our experiments, multi-talker ASR models trained with the proposed method show consistent improvement on the word error rates across multiple datasets.

Neural transducer is now the most popular end-to-end model for speech recognition, due to its naturally streaming ability. However, it is challenging to adapt it with text-only data. Factorized neural transducer (FNT) model was proposed to mitigate this problem. The improved adaptation ability of FNT on text-only adaptation data came at the cost of lowered accuracy compared to the standard neural transducer model. We propose several methods to improve the performance of the FNT model. They are: adding CTC criterion during training, adding KL divergence loss during adaptation, using a pre-trained language model to seed the vocabulary predictor, and an efficient adaptation approach by interpolating the vocabulary predictor with the n-gram language model. A combination of these approaches results in a relative word-error-rate reduction of 9.48% from the standard FNT model. Furthermore, n-gram interpolation with the vocabulary predictor improves the adaptation speed hugely with satisfactory adaptation performance.

Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

The rapid development of single-modal pretraining has prompted researchers to pay more attention to cross-modal pre-training methods.In this paper, we propose a unified-modal speech-unit-text pre-training model, SpeechUT, to connect the representations of a speech encoder and a text decoder with a shared unit encoder.Leveraging hidden-unit as an interface to align speech and text, we can decompose the speech-to-text model into a speech-to-unit model and a unit-to-text model, which can be jointly pre-trained with unpaired speech and text data respectively.Our proposed SpeechUT is fine-tuned and evaluated on automatic speech recognition (ASR) and speech translation (ST) tasks.Experimental results show that SpeechUT gets substantial improvements over strong baselines, and achieves state-of-the-art performance on both the Lib-riSpeech ASR and MuST-C ST tasks.To better understand the proposed SpeechUT, detailed analyses are conducted.The code and pretrained models are available at https://aka. ms/SpeechUT.

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.

Self-supervised learning (SSL) is a long-standing goal for speech processing, since it utilizes large-scale unlabeled data and avoids extensive human labeling. Recent years have witnessed great successes in applying self-supervised learning in speech recognition, while limited exploration was attempted in applying SSL for modeling speaker characteristics. In this paper, we aim to improve the existing SSL framework for speaker representation learning. Two methods are introduced for enhancing the unsupervised speaker information extraction. First, we apply multi-task learning to the current SSL framework, where we integrate utterance-wise contrastive loss with the SSL objective function. Second, for better speaker discrimination, we propose an utterance mixing strategy for data augmentation, where additional overlapped utterances are created unsupervisely and incorporated during training. We integrate the proposed methods into the HuBERT framework. Experiment results on the SUPERB benchmark show that the proposed system achieves state-of-the-art performance in universal representation learning, especially for speaker identification oriented tasks. An ablation study is performed verifying the efficacy of each proposed method. Finally, we scale up the training dataset to 94 thousand hours of public audio data and achieve further performance improvement in all SUPERB tasks.

Streaming end-to-end multi-talker speech recognition aims at transcribing the overlapped speech from conversations or meetings with an all-neural model in a streaming fashion, which is fundamentally different from a modular-based approach that usually cascades the speech separation and the speech recognition models trained independently. Previously, we proposed the Streaming Unmixing and Recognition Transducer (SURT) model based on recurrent neural network transducer (RNN-T) for this problem and presented promising results. However, for real applications, the speech recognition system is also required to determine the times-tamp when a speaker finishes speaking for prompt system response. This problem, known as endpoint (EP) detection, has not been studied previously for multi-talker end-to-end models. In this work, we address the EP detection problem in the SURT framework by introducing an end-of-sentence token as an output unit, following the practice of single-talker end-to-end models. Furthermore, we also present a latency penalty approach that can significantly cut down the EP detection latency. Our experimental results based on the 2-speaker LibrispeechMix dataset show that the SURT model can achieve promising EP detection without significantly degradation of the recognition accuracy.

Streaming recognition of multi-talker conversations has so far been evaluated only for 2-speaker single-turn sessions. In this paper, we investigate it for multi-turn meetings containing multiple speakers using the Streaming Unmixing and Recognition Transducer (SURT) model, and show that naively extending the single-turn model to this harder setting incurs a performance penalty. As a solution, we propose the dual-path (DP) modeling strategy first used for time-domain speech separation. We experiment with LSTM and Transformer based DP models, and show that they improve word error rate (WER) performance while yielding faster convergence. We also explore training strategies such as chunk width randomization and curriculum learning for these models, and demonstrate their importance through ablation studies. Finally, we evaluate our models on the LibriCSS meeting data, where they perform competitively with offline separation-based methods.

The goal of self-supervised learning (SSL) for automatic speech recognition (ASR) is to learn good speech representations from a large amount of unlabeled speech for the downstream ASR task. However, most SSL frameworks do not consider noise robustness which is crucial for real-world applications. In this paper we propose wav2vec-Switch, a method to encode noise robustness into contextualized representations of speech via contrastive learning. Specifically, we feed original-noisy speech pairs simultaneously into the wav2vec 2.0 network. In addition to the existing contrastive learning task, we switch the quantized representations of the original and noisy speech as additional prediction targets of each other. By doing this, it enforces the network to have consistent predictions for the original and noisy speech, thus allows to learn contextualized representation with noise robustness. Our experiments on synthe-sized and real noisy data show the effectiveness of our method: it achieves 2.9–4.9% relative word error rate (WER) reduction on the synthesized noisy LibriSpeech data without deterioration on the original data, and 5.7% on CHiME-4 real 1-channel noisy data compared to a data augmentation baseline even with a strong language model for decoding. Our results on CHiME-4 can match or even surpass those with well-designed speech enhancement components.

Noise robustness is essential for deploying automatic speech recognition (ASR) systems in real-world environments. One way to reduce the effect of noise interference is to employ a preprocessing module that conducts speech enhancement, and then feed the enhanced speech to an ASR backend. In this work, instead of suppressing background noise with a conventional cascaded pipeline, we employ a noise-robust representation learned by a refined self-supervised framework for noisy speech recognition. We propose to combine a reconstruction module with contrastive learning and perform multi-task continual pre-training on noisy data. The reconstruction module is used for auxiliary learning to improve the noise robustness of the learned representation and thus is not required during inference. Experiments demonstrate the effectiveness of our proposed method. Our model substantially reduces the word error rate (WER) for the synthesized noisy LibriSpeech test sets, and yields around 4.1/7.5% WER reduction on noisy clean/other test sets compared to data augmentation. For the real-world noisy speech from the CHiME-4 challenge (1-channel track), we have obtained the state of the art ASR performance without any denoising front-end. Moreover, we achieve comparable performance to the best supervised approach reported with only 16% of labeled data.

Recently, pioneer work finds that self-supervised pre-training methods can improve multiple downstream speech tasks, because the model utilizes bottom layers to learn speaker-related information and top layers to encode content-related information. Since the network capacity is limited, we believe the speech recognition performance could be further improved if the model is dedicated to audio content information learning. To this end, we propose Intermediate Layer Supervision for Self-Supervised Learning (ILS-SSL), which forces the model to concentrate on content information as much as possible by adding an additional SSL loss on the intermediate layers. Experiments on LibriSpeech test-other set show that our method outperforms HuBERT significantly, which achieves a 23.5%/11.6% relative word error rate reduction in the w/o language model setting for Base/Large models. Detailed analysis shows the bottom layers of our model have a better correlation with phonetic units, which is consistent with our intuition and explains the success of our method for ASR. We will release our code and model at https://github.com/microsoft/UniSpeech.

Hybrid and end-to-end (E2E) systems have their individual advantages, with different error patterns in the speech recognition results. By jointly modeling audio and text, the E2E model performs better in matched scenarios and scales well with a large amount of paired audio-text training data. The modularized hybrid model is easier for customization, and better to make use of a massive amount of unpaired text data. This paper proposes a two-pass hybrid and E2E cascading (HEC) framework to combine the hybrid and E2E model in order to take advantage of both sides, with hybrid in the first pass and E2E in the second pass. We show that the proposed system achieves 8-10% relative word error rate reduction with respect to each individual system. More importantly, compared with the pure E2E system, we show the proposed system has the potential to keep the advantages of hybrid system, e.g., customization and segmentation capabilities. We also show the second pass E2E model in HEC is robust with respect to the change in the first pass hybrid model.

While permutation invariant training (PIT) based continuous speech separation (CSS) significantly improves the conversation transcription accuracy, it often suffers from speech leakages and failures in separation at "hot spot" regions because it has a fixed number of output channels. In this paper, we propose to apply recurrent selective attention network (RSAN) to CSS, which generates a variable number of output channels based on active speaker counting. In addition, we propose a novel block-wise dependency extension of RSAN by introducing dependencies between adjacent processing blocks in the CSS framework. It enables the network to utilize the separation results from the previous blocks to facilitate the current block processing. Experimental results on the LibriCSS dataset show that the RSAN-based CSS (RSAN-CSS) network consistently improves the speech recognition accuracy over PIT-based models. The proposed block-wise dependency modeling further boosts the performance of RSAN-CSS.

Multilingual automatic speech recognition models have shown great promise in recent years because of the simple model training and deployment process. Conventional methods either train a universal multilingual model without taking any language information or with a 1-hot language ID (LID) vector to guide the recognition of the target language. In practice, a multilingual user can be prompted to preselect several languages he/she can speak. The multilingual model without LID cannot well utilize the language information set by the user while the multilingual model with 1-hot LID can only handle one pre-selected language. In this paper, we propose a novel configurable multilingual model (CMM) which is trained only once but can be configured as different models based on users’ choices by extracting language-specific modules together with a universal module from the trained CMM. Particularly, a single CMM can be deployed to any user scenario where the users can pre-select any combination of languages. Trained with 75K hours of transcribed anonymized Microsoft multilingual data and evaluated with 10-language test sets, the proposed CMM improves from the universal multilingual model by 26.0%, 16.9%, and 10.4% relative word error reduction when the user selects 1, 2, or 3 languages, respectively.

This paper studies a novel pre-training technique with unpaired speech data, Speech2C, for encoder-decoder based automatic speech recognition (ASR).Within a multi-task learning framework, we introduce two pre-training tasks for the encoderdecoder network using acoustic units, i.e., pseudo codes, derived from an offline clustering model.One is to predict the pseudo codes via masked language modeling in encoder output, like HuBERT model, while the other lets the decoder learn to reconstruct pseudo codes autoregressively instead of generating textual scripts.In this way, the decoder learns to reconstruct original speech information with codes before learning to generate correct text.Comprehensive experiments on the LibriSpeech corpus show that the proposed Speech2C can relatively reduce the word error rate (WER) by 19.2% over the method without decoder pre-training, and also outperforms significantly the state-of-the-art wav2vec 2.0 and Hu-BERT on fine-tuning subsets of 10h and 100h.We release our code and model at https://github.com/microsoft/SpeechT5/tree/main/Speech2C.

Recently, self-supervised learning (SSL) has demonstrated strong performance in speaker recognition, even if the pretraining objective is designed for speech recognition.In this paper, we study which factor leads to the success of selfsupervised learning on speaker-related tasks, e.g.speaker verification (SV), through a series of carefully designed experiments.Our empirical results on the Voxceleb-1 dataset suggest that the benefit of SSL to SV task is from a combination of mask speech prediction loss, data scale, and model size, while the SSL quantizer has a minor impact.We further employ the integrated gradients attribution method and loss landscape visualization to understand the effectiveness of self-supervised learning for speaker recognition performance.

This paper presents a streaming speaker-attributed automatic speech recognition (SA-ASR) model that can recognize "who spoke what" with low latency even when multiple people are speaking simultaneously.Our model is based on token-level serialized output training (t-SOT) which was recently proposed to transcribe multi-talker speech in a streaming fashion.To further recognize speaker identities, we propose an encoderdecoder based speaker embedding extractor that can estimate a speaker representation for each recognized token not only from non-overlapping speech but also from overlapping speech.The proposed speaker embedding, named t-vector, is extracted synchronously with the t-SOT ASR model, enabling joint execution of speaker identification (SID) or speaker diarization (SD) with the multi-talker transcription with low latency.We evaluate the proposed model for a joint task of ASR and SID/SD by using LibriSpeechMix and LibriCSS corpora.The proposed model achieves substantially better accuracy than a prior streaming model and shows comparable or sometimes even superior results to the state-of-the-art offline SA-ASR model.

This paper proposes a token-level serialized output training (t-SOT), a novel framework for streaming multi-talker automatic speech recognition (ASR). Unlike existing streaming multi-talker ASR models using multiple output branches, the t-SOT model has only a single output branch that generates recognition tokens (e.g., words, subwords) of multiple speakers in chronological order based on their emission times. A special token that indicates the change of ``virtual'' output channels is introduced to keep track of the overlapping utterances. Compared to the prior streaming multi-talker ASR models, the t-SOT model has the advantages of less inference cost and a simpler model architecture. Moreover, in our experiments with LibriSpeechMix and LibriCSS datasets, the t-SOT-based transformer transducer model achieves the state-of-the-art word error rates by a significant margin to the prior results. For non-overlapping speech, the t-SOT model is on par with a single-talker ASR model in terms of both accuracy and computational cost, opening the door for deploying one model for both single- and multi-talker scenarios.

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, masked prediction pre-training has seen remarkable progress in self-supervised learning (SSL) for speech recognition.It usually requires a codebook obtained in an unsupervised way, making it less accurate and difficult to interpret.We propose two supervision-guided codebook generation approaches to improve automatic speech recognition (ASR) performance and also the pre-training efficiency, either through decoding with a hybrid ASR system to generate phoneme-level alignments (named PBERT), or performing clustering on the supervised speech features extracted from an end-to-end CTC model (named CTC clustering).Both the hybrid and CTC models are trained on the same small amount of labeled speech as used in fine-tuning.Experiments demonstrate significant superiority of our methods to various SSL and self-training baselines, with up to 17.0% relative WER reduction.Our pre-trained models also show good transferability in a non-ASR speech task.

Neural transducers have been widely used in automatic speech recognition (ASR). In this paper, we introduce it to streaming end-to-end speech translation (ST), which aims to convert audio signals to texts in other languages directly. Compared with cascaded ST that performs ASR followed by text-based machine translation (MT), the proposed Transformer transducer (TT)-based ST model drastically reduces inference latency, exploits speech information, and avoids error propagation from ASR to MT. To improve the modeling capacity, we propose attention pooling for the joint network in TT. In addition, we extend TT-based ST to multilingual ST, which generates texts of multiple languages at the same time. Experimental results on a large-scale 50 thousand (K) hours pseudo-labeled training set show that TT-based ST not only significantly reduces inference time but also outperforms non-streaming cascaded ST for English-German translation.

Multi-talker conversational speech processing has drawn many interests for various applications such as meeting transcription.Speech separation is often required to handle overlapped speech that is commonly observed in conversation.Although the original utterancelevel permutation invariant training-based continuous speech separation approach has proven to be effective in various conditions, it lacks the ability to leverage the long-span relationship of utterances and is computationally inefficient due to the highly overlapped sliding windows.To overcome these drawbacks, we propose a novel training scheme named Group-PIT, which allows direct training of the speech separation models on the long-form speech with a low computational cost for label assignment.Two different speech separation approaches with Group-PIT are explored, including direct long-span speech separation and short-span speech separation with long-span tracking.The experiments on the simulated meeting-style data demonstrate the effectiveness of our proposed approaches, especially in dealing with a very long speech input.

Automatic Speech Recognition (ASR) systems typically yield output in lexical form. However, humans prefer a written form output. To bridge this gap, ASR systems usually employ Inverse Text Normalization (ITN). In previous works, Weighted Finite State Transducers (WFST) have been employed to do ITN. WFSTs are nicely suited to this task but their size and run-time costs can make deployment on embedded applications challenging. In this paper, we describe the development of an on-device ITN system that is streaming, lightweight & accurate. At the core of our system is a streaming transformer tagger, that tags lexical tokens from ASR. The tag informs which ITN category might be applied, if at all. Following that, we apply an ITN-category-specific WFST, only on the tagged text, to reliably perform the ITN conversion. We show that the proposed ITN solution performs equivalent to strong base-lines, while being significantly smaller in size and retaining customization capabilities.

Contextual biasing is an important and challenging task for end-to-end automatic speech recognition (ASR) systems, which aims to achieve better recognition performance by biasing the ASR system to particular context phrases such as person names, music list, proper nouns, etc. Existing methods mainly include contextual LM biasing and adding bias encoder into end-to-end ASR models. In this work, we introduce a novel approach to do contextual biasing by adding a contextual spelling correction model on top of the end-to-end ASR system. We incorporate contextual information into a sequence-to-sequence spelling correction model with a shared context encoder. The proposed model includes two different mechanisms: autoregressive (AR) and non-autoregressive (NAR). We also propose filtering algorithms to handle large-size context lists, and performance balancing mechanisms to control the biasing degree of the model. The proposed model is a general biasing solution which is domain-insensitive and can be adopted in different scenarios. Experiments show that the proposed method achieves as much as 51% relative word error rate (WER) reduction over ASR system and outperforms traditional biasing methods. Compared to the AR solution, the NAR model reduces model size by 43.2% and speeds up inference by 2.1 times.

In this paper, several works are proposed to address practi-cal challenges for deploying RNN Transducer (RNN-T) based speech recognition systems. These challenges are adapting a well-trained RNN-T model to a new domain without col-lecting the audio data, obtaining time stamps and confidence scores at word level. We solve the first challenge with a splicing data method which concatenates the speech segments ex-tracted from the source domain data. To get time stamps, a phone prediction branch is added to the RNN-T model by sharing the encoder for the purpose of forced alignment. Fi-nally, we obtain word level confidence scores by utilizing sev-eral types of features calculated during decoding and from a confusion network. Evaluated with Microsoft production data, the splicing data adaptation method improves the base-line and adaptation with the text to speech method by 58.03% and 15.25% relative word error rate reduction, respectively. The proposed time stamping method can get less than 50 mil-lisecond word timing difference from the ground truth align-ment on average while maintaining the recognition accuracy. We also obtain high confidence annotation performance with limited computation cost.

With its strong modeling capacity that comes from a multi-head and multi-layer structure, Transformer is a very powerful model for learning a sequential representation and has been successfully applied to speech separation recently. However, multi-channel speech separation sometimes does not necessarily need such a heavy structure for all time frames especially when the cross-talker challenge happens only occasionally. For example, in conversation scenarios, most regions contain only a single active speaker, where the separation task downgrades to a single speaker enhancement problem. It turns out that using a very deep network structure for dealing with signals with a low overlap ratio not only negatively affects the inference efficiency but also hurts the separation performance. To deal with this problem, we propose an early exit mechanism, which enables the Transformer model to handle different cases with adaptive depth. Experimental results indicate that not only does the early exit mechanism accelerate the inference, but it also improves the accuracy.

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.

Continuous speech separation was recently proposed to deal with the overlapped speech in natural conversations. While it was shown to significantly improve the speech recognition performance for multichannel conversation transcription, its effectiveness has yet to be proven for a single-channel recording scenario. This paper examines the use of Conformer architecture in lieu of recurrent neural networks for the separation model. Conformer allows the separation model to efficiently capture both local and global context information, which is helpful for speech separation. Experimental results using the LibriCSS dataset show that the Conformer separation model achieves the state of the art results for both single-channel and multi-channel settings. Results for real meeting recordings are also presented, showing significant performance gains in both word error rate (WER) and speaker-attributed WER.

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.

Hypothesis-level combination between multiple models can often yield gains in speech recognition. However, all models in the ensemble are usually restricted to use the same audio segmentation times. This paper proposes to generalise hypothesis-level combination, allowing the use of different audio segmentation times between the models, by splitting and re-joining the hypothesised N-best lists in time. A hypothesis tree method is also proposed to distribute hypothesis posteriors among the constituent words, to facilitate such splitting when per-word scores are not available. The approach is assessed on a Microsoft meeting transcription task, by performing combination between a streaming first-pass recognition and an offline second-pass recognition. The experimental results show that the proposed approach can yield gains when combining over different segmentation times. Furthermore, the results also show that a combination between a hybrid model and an end-to-end neural network model yields a greater improvement than a combination between two hybrid models.

This paper describes the Microsoft speaker diarization system for monaural multi-talker recordings in the wild, evaluated at the diarization track of the VoxCeleb Speaker Recognition Challenge (VoxSRC) 2020. We will first explain our system design to address issues in handling real multi-talker recordings. We then present the details of the components, which include Res2Net-based speaker embedding extractor, conformer-based continuous speech separation with leakage filtering, and a modified DOVER (short for Diarization Output Voting Error Reduction) method for system fusion. We evaluate the systems with the data set provided by VoxSRC challenge 2020, which contains real-life multi-talker audio collected from YouTube. Our best system achieves 3.71% and 6.23% of the diarization error rate (DER) on development set and evaluation set, respectively, being ranked the 1st at the diarization track of the challenge.

In multi-talker scenarios such as meetings and conversations, speech processing systems are usually required to transcribe the audio as well as identify the speakers for downstream applications. Since overlapped speech is common in this case, conventional approaches usually address this problem in a cascaded fashion that involves speech separation, speech recognition and speaker identification that are trained independently. In this paper, we propose Streaming Unmixing, Recognition and Identification Transducer (SURIT) -- a new framework that deals with this problem in an end-to-end streaming fashion. SURIT employs the recurrent neural network transducer (RNN-T) as the backbone for both speech recognition and speaker identification. We validate our idea on the LibrispeechMix dataset -- a multi-talker dataset derived from Librispeech, and present encouraging results.

Hybrid Autoregressive Transducer (HAT) is a recently proposed end-to-end acoustic model that extends the standard Recurrent Neural Network Transducer (RNN-T) for the purpose of the external language model (LM) fusion.In HAT, the blank probability and the label probability are estimated using two separate probability distributions, which provides a more accurate solution for internal LM score estimation, and thus works better when combining with an external LM.Previous work mainly focuses on HAT model training with the negative log-likelihood loss, while in this paper, we study the minimum word error rate (MWER) training of HAT -a criterion that is closer to the evaluation metric for speech recognition, and has been successfully applied to other types of end-to-end models such as sequenceto-sequence (S2S) and RNN-T models.From experiments with around 30,000 hours of training data, we show that MWER training can improve the accuracy of HAT models, while at the same time, improving the robustness of the model against the decoding hyper-parameters such as length normalization and decoding beam during inference.

Transformer has been successfully applied to speech separation recently with its strong long-dependency modeling capacity using a self-attention mechanism. However, Transformer tends to have heavy run-time costs due to the deep encoder layers, which hinders its deployment on edge devices. A small Transformer model with fewer encoder layers is preferred for computational efficiency, but it is prone to performance degradation. In this paper, an ultra fast speech separation Transformer model is proposed to achieve both better performance and efficiency with teacher student learning (T-S learning). We introduce layer-wise T-S learning and objective shifting mechanisms to guide the small student model to learn intermediate representations from the large teacher model. Compared with the small Transformer model trained from scratch, the proposed T-S learning method reduces the word error rate (WER) by more than 5% for both multi-channel and single-channel speech separation on LibriCSS dataset. Utilizing more unlabeled speech data, our ultra fast speech separation models achieve more than 10% relative WER reduction.

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.

It's challenging to customize transducer-based automatic speech recognition (ASR) system with context information which is dynamic and unavailable during model training. In this work, we introduce a light-weight contextual spelling correction model to correct context-related recognition errors in transducer-based ASR systems. We incorporate the context information into the spelling correction model with a shared context encoder and use a filtering algorithm to handle large-size context lists. Experiments show that the model improves baseline ASR model performance with about 50% relative word error rate reduction, which also significantly outperforms the baseline method such as contextual LM biasing. The model also shows excellent performance for out-of-vocabulary terms not seen during training.

Speech separation has been successfully applied as a frontend processing module of conversation transcription systems thanks to its ability to handle overlapped speech and its flexibility to combine with downstream tasks such as automatic speech recognition (ASR). However, a speech separation model often introduces target speech distortion, resulting in a sub-optimum word error rate (WER). In this paper, we describe our efforts to improve the performance of a single channel speech separation system. Specifically, we investigate a two-stage training scheme that firstly applies a feature level optimization criterion for pretraining, followed by an ASR-oriented optimization criterion using an end-to-end (E2E) speech recognition model. Meanwhile, to keep the model light-weight, we introduce a modified teacher-student learning technique for model compression. By combining those approaches, we achieve a absolute average WER improvement of 2.70% and 0.77% using models with less than 10M parameters compared with the previous state-of-the-art results on the LibriCSS dataset for utterance-wise evaluation and continuous evaluation, respectively

In this study, we try to address the problem of leveraging visual signals to improve Automatic Speech Recognition (ASR), also known as visual context-aware ASR (VC-ASR). We explore novel VC-ASR approaches to leverage video and text representations extracted by a self-supervised pre-trained text-video embedding model. Firstly, we propose a multi-stream attention architecture to leverage signals from both audio and video modalities. This architecture consists of separate encoders for the two modalities and a single decoder that at-tends over them. We show that this architecture is better than fusing modalities at the signal level. Additionally, we also explore lever-aging the visual information in a second pass model, which has also been referred to as a `deliberation model'. The deliberation model accepts audio representations and text hypotheses from the first pass ASR and combines them with a visual stream for an improved visual context-aware recognition. The proposed deliberation scheme can work on top of any well trained ASR and also enabled us to leverage the pre-trained text model to ground the hypotheses with the visual features. Our experiments on HOW2 dataset show that multi-stream and deliberation architectures are very effective at the VC-ASR task. We evaluate the proposed models for two scenarios; clean audio stream and distorted audio in which we mask out some specific words in the audio. The deliberation model outperforms the multi-stream model and achieves a relative WER improvement of 6% and 8.7% for the clean and masked data, respectively, compared to an audio-only model. The deliberation model also improves re-covering the masked words by 59% relative.

Continuous speech separation (CSS) is an arising task in speech separation aiming at separating overlap-free targets from a long, partially-overlapped recording. A straightforward extension of previously proposed sentence-level separation models to this task is to segment the long recording into fixed-length blocks and perform separation on them independently. However, such simple extension does not fully address the cross-block dependencies and the separation performance may not be satisfactory. In this paper, we focus on how the block-level separation performance can be improved by exploring methods to utilize the cross-block information. Based on the recently proposed dual-path RNN (DPRNN) architecture, we investigate how DPRNN can help the block-level separation by the interleaved intra- and inter-block modules. Experiment results show that DPRNN is able to significantly outperform the baseline block-level model in both offline and block-online configurations under certain settings.

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.

Multi-speaker speech recognition of unsegmented recordings has diverse applications such as meeting transcription and automatic subtitle generation. With technical advances in systems dealing with speech separation, speaker diarization, and automatic speech recognition (ASR) in the last decade, it has become possible to build pipelines that achieve reasonable error rates on this task. In this paper, we propose an end-to-end modular system for the LibriCSS meeting data, which combines independently trained separation, diarization, and recognition components, in that order. We study the effect of different state-of-the-art methods at each stage of the pipeline, and report results using task-specific metrics like SDR and DER, as well as downstream WER. Experiments indicate that the problem of overlapping speech for diarization and ASR can be effectively mitigated with the presence of a well-trained separation module. Our best system achieves a speaker-attributed WER of 12.7%, which is close to that of a non-overlapping ASR.

End-to-end multi-talker speech recognition is an emerging research trend in the speech community due to its vast potential in applications such as conversation and meeting transcriptions. To the best of our knowledge, all existing research works are constrained in the offline scenario. In this work, we propose the Streaming Unmixing and Recognition Transducer (SURT) for end-to-end multi-talker speech recognition. Our model employs the Recurrent Neural Network Transducer (RNN-T) as the backbone that can meet various latency constraints. We study two different model architectures that are based on a speaker-differentiator encoder and a mask encoder respectively. To train this model, we investigate the widely used Permutation Invariant Training (PIT) approach and the Heuristic Error Assignment Training (HEAT) approach. Based on experiments on the publicly available LibriSpeechMix dataset, we show that HEAT can achieve better accuracy compared with PIT, and the SURT model with 150 milliseconds algorithmic latency constraint compares favorably with the offline sequence-to-sequence based baseline model in terms of accuracy.

We propose speaker separation using speaker inventories and estimated speech (SSUSIES), a framework leveraging speaker profiles and estimated speech for speaker separation. SSUSIES contains two methods, speaker separation using speaker inventories (SSUSI) and speaker separation using estimated speech (SSUES). SSUSI performs speaker separation with the help of speaker inventory. By combining the advantages of permutation invariant training (PIT) and speech extraction, SSUSI significantly outperforms conventional approaches. SSUES is a widely applicable technique that can substantially improve speaker separation performance using the output of first-pass separation. We evaluate the models on both speaker separation and speech recognition metrics.

We propose to use the personalized speech synthesis and the neural language generator to synthesize content relevant personalized speech for rapid speaker adaptation. It has two distinct aspects: First, it relieves the general data sparsity issue in rapid adaptation via making use of additional synthesized personalized speech; Second, it circumvents the obstacle of the explicit labeling error in unsupervised adaptation by converting it to pseudo-supervised adaptation. In this setup, the labeling error is implicitly rendered as less damaging speech distortion in the personalized synthesized speech. This results in significant performance breakthrough in the rapid unsupervised speaker adaptation. We apply the proposed methodology to a speaker adaptation task in a state-of-art speech transcription system. With 1 minute (min) adaptation data, our proposed approach yields 9.19 % or 5.98 % relative word error rate (WER) reduction for the supervised and the unsupervised adaptation, comparing to the negligible gain when adapting only with 1 min original speech. With 10 min adaptation data, it yields 12.53 % or 7.89 % relative WER reduction, doubling the gain of the baseline adaptation. The proposed approach is particularly suitable for unsupervised adaptation.

This paper describes a dataset and protocols for evaluating continuous speech separation algorithms. Most prior speech separation studies use pre-segmented audio signals, which are typically generated by mixing speech utterances on computers so that they fully overlap. Also, the separation algorithms have often been evaluated based on signal-based metrics such as signal-to-distortion ratio. However, in natural conversations, speech signals are continuous and contain both overlapped and overlap-free regions. In addition, the signal-based metrics only have weak correlation with automatic speech recognition (ASR) accuracy. Not only does this make it hard to assess the practical relevance of the tested algorithms, it also hinders researchers from developing systems that can be readily applied to real scenarios. In this paper, we define continuous speech separation (CSS) as a task of generating a set of non-overlapped speech signals from a continuous audio stream that contains multiple utterances that are partially overlapped by a varying degree. A new real recording dataset, called LibriCSS, is derived from LibriSpeech by concatenating the corpus utterances to simulate conversations and capturing the audio replays with far-field microphones. A Kaldi-based ASR evaluation protocol is established by using a well-trained multi-conditional acoustic model. A recently proposed speaker-independent CSS algorithm is investigated by using LibriCSS. The dataset and evaluation scripts are made available to facilitate the research in this direction1.

Recently, the recurrent neural network transducer (RNN-T) architecture has become an emerging trend in end-to-end automatic speech recognition research due to its advantages of being capable for online streaming speech recognition. However, RNN-T training is made difficult by the huge memory requirements, and complicated neural structure. A common solution to ease the RNN-T training is to employ connectionist temporal classification (CTC) model along with RNN language model (RNNLM) to initialize the RNN-T parameters. In this work, we conversely leverage external alignments to seed the RNN-T model. Two different pre-training solutions are explored, referred to as encoder pre-training, and whole-network pre-training respectively. Evaluated on Microsoft 65,000 hours anonymized production data with personally identifiable information removed, our proposed methods can obtain significant improvement. In particular, the encoder pre-training solution achieved a 10% and a 8% relative word error rate reduction when compared with random initialization and the widely used CTC+RNNLM initialization strategy, respectively. Our solutions also significantly reduce the RNN-T model latency from the baseline.

Recently, a few novel streaming attention-based sequence-to-sequence (S2S) models have been proposed to perform online speech recognition with linear-time decoding complexity. However, in these models, the decisions to generate tokens are delayed compared to the actual acoustic boundaries since their unidirectional encoders lack future information. This leads to an inevitable latency during inference. To alleviate this issue and reduce latency, we propose several strategies during training by leveraging external hard alignments extracted from the hybrid model. We investigate to utilize the alignments in both the encoder and the decoder. On the encoder side, (1) multi-task learning and (2) pre-training with the framewise classification task are studied. On the decoder side, we (3) remove inappropriate alignment paths beyond an acceptable latency during the alignment marginalization, and (4) directly min-imize the differentiable expected latency loss. Experiments on the Cortana voice search task demonstrate that our proposed methods can significantly reduce the latency, and even improve the recognition accuracy in certain cases on the decoder side. We also present some analysis to understand the behaviors of streaming S2S models.

While the community keeps promoting end-to-end models over conventional hybrid models, which usually are long short-term memory (LSTM) models trained with a cross entropy criterion followed by a sequence discriminative training criterion, we argue that such conventional hybrid models can still be significantly improved. In this paper, we detail our recent efforts to improve conventional hybrid LSTM acoustic models for high-accuracy and low-latency automatic speech recognition. To achieve high accuracy, we use a contextual layer trajectory LSTM (cltLSTM), which decouples the temporal modeling and target classification tasks, and incorporates future context frames to get more information for accurate acoustic modeling. We further improve the training strategy with sequence-level teacher-student learning. To obtain low latency, we design a two-head cltLSTM, in which one head has zero latency and the other head has a small latency, compared to an LSTM. When trained with Microsoft's 65 thousand hours of anonymized training data and evaluated with test sets with 1.8 million words, the proposed two-head cltLSTM model with the proposed training strategy yields a 28.2% relative WER reduction over the conventional LSTM acoustic model, with a similar perceived latency.

We propose a novel neural label embedding (NLE) scheme for the domain adaptation of a deep neural network (DNN) acoustic model with unpaired data samples from source and target domains. With NLE method, we distill the knowledge from a powerful source-domain DNN into a dictionary of label embeddings, or l-vectors, one for each senone class. Each l-vector is a representation of the senone-specific output distributions of the source-domain DNN and is learned to minimize the average L2, Kullback-Leibler (KL) or symmetric KL distance to the output vectors with the same label through simple averaging or standard back-propagation. During adaptation, the l-vectors serve as the soft targets to train the target-domain model with cross-entropy loss. Without parallel data constraint as in the teacher-student learning, NLE is specially suited for the situation where the paired target-domain data cannot be simulated from the source-domain data. We adapt a 6400 hours multi-conditional US English acoustic model to each of the 9 accented English (80 to 830 hours) and kids' speech (80 hours). NLE achieves up to 14.1% relative word error rate reduction over direct re-training with one-hot labels.

Attention-based encoder-decoder model has achieved impressive results for both automatic speech recognition (ASR) and text-to-speech (TTS) tasks.This approach takes advantage of the memorization capacity of neural networks to learn the mapping from the input sequence to the output sequence from scratch, without the assumption of prior knowledge such as the alignments.However, this model is prone to overfitting, especially when the amount of training data is limited.Inspired by SpecAugment and BERT, in this paper, we propose a semantic mask based regularization for training such kind of end-toend (E2E) model.The idea is to mask the input features corresponding to a particular output token, e.g., a word or a wordpiece, in order to encourage the model to fill the token based on the contextual information.While this approach is applicable to the encoder-decoder framework with any type of neural network architecture, we study the transformer-based model for ASR in this work.We perform experiments on Librispeech 960h and TedLium2 data sets, and achieve the state-of-the-art performance on the test set in the scope of E2E models.

The attention-based Transformer model has achieved promising results for speech recognition (SR) in the offline mode.However, in the streaming mode, the Transformer model usually incurs significant latency to maintain its recognition accuracy when applying a fixed-length look-ahead window in each encoder layer.In this paper, we propose a novel low-latency streaming approach for Transformer models, which consists of a scout network and a recognition network.The scout network detects the whole word boundary without seeing any future frames, while the recognition network predicts the next subword by utilizing the information from all the frames before the predicted boundary.Our model achieves the best performance (2.7/6.4WER) with only 639 ms latency on the test-clean and test-other data sets of Librispeech.

Rapid unsupervised speaker adaptation in an E2E system posits us new challenges due to its end-to-end unified structure in addition to its intrinsic difficulty of data sparsity and imperfect label [1]. Previously we proposed utilizing the content relevant personalized speech synthesis for rapid speaker adaptation and achieved significant performance breakthrough in a hybrid system [2]. In this paper, we answer the following two questions: First, how to effectively perform rapid speaker adaptation in an RNN-T. Second, whether our previously proposed approach is still beneficial for the RNN-T and what are the modification and distinct observations. We apply the proposed methodology to a speaker adaptation task in a state-of-art presentation transcription RNN-T system. In the 1 min setup, it yields 11.58 % or 7.95 % relative word error rate (WER) reduction for the sup/unsup adaptation, comparing to the negligible gain when adapting with 1 min source speech. In the 10 min setup, it yields 15.71 % or 8.00 % relative WER reduction, doubling the gain of the source speech adaptation. We further apply various data filtering techniques and significantly bridge the gap between sup/unsup adaptation.

Transfer learning (TL) is widely used in conventional hybrid automatic speech recognition (ASR) system, to transfer the knowledge from source to target language. TL can be applied to end-to-end (E2E) ASR system such as recurrent neural network transducer (RNN-T) models, by initializing the encoder and/or prediction network of the target language with the pre-trained models from source language. In the hybrid ASR system, transfer learning is typically done by initializing the target language acoustic model (AM) with source language AM. Several transfer learning strategies exist in the case of the RNN-T framework, depending upon the choice of the initialization model for encoder and prediction networks. This paper presents a comparative study of four different TL methods for RNN-T framework. We show ~10%-17% relative word error rate reduction with different TL methods over randomly initialized RNN-T model. We also study the impact of TL with varying amount of training data ranging from 50 hours to 1000 hours and show the efficacy of TL for languages with a very small amount of training data.

In this work, we develop new self-learning techniques with an attention-based sequence-to-sequence (seq2seq) model for automatic speech recognition (ASR). For untranscribed speech data, the hypothesis from an ASR system must be used as a label. However, the imperfect ASR result makes unsupervised learning difficult to consistently improve recognition performance especially in the case that multiple powerful teacher models are unavailable. In contrast to conventional unsupervised learning approaches, we adopt the \emph{multi-task learning} (MTL) framework where the $n$-th best ASR hypothesis is used as the label of each task. The seq2seq network is updated through the MTL framework so as to find the common representation that can cover multiple hypotheses. By doing so, the effect of the \emph{hard-decision} errors can be alleviated. We first demonstrate the effectiveness of our self-learning methods through ASR experiments in an accent adaptation task between the US and British English speech. Our experiment results show that our method can reduce the WER on the British speech data from 14.55\% to 10.36\% compared to the baseline model trained with the US English data only. Moreover, we investigate the effect of our proposed methods in a federated learning scenario.

Recently, there has been a strong push to transition from hybrid models to end-to-end (E2E) models for automatic speech recognition.Currently, there are three promising E2E methods: recurrent neural network transducer (RNN-T), RNN attentionbased encoder-decoder (AED), and Transformer-AED.In this study, we conduct an empirical comparison of RNN-T, RNN-AED, and Transformer-AED models, in both non-streaming and streaming modes.We use 65 thousand hours of Microsoft anonymized training data to train these models.As E2E models are more data hungry, it is better to compare their effectiveness with large amount of training data.To the best of our knowledge, no such comprehensive study has been conducted yet.We show that although AED models are stronger than RNN-T in the non-streaming mode, RNN-T is very competitive in streaming mode if its encoder can be properly initialized.Among all three E2E models, transformer-AED achieved the best accuracy in both streaming and non-streaming mode.We show that both streaming RNN-T and transformer-AED models can obtain better accuracy than a highly-optimized hybrid model.

Because of its streaming nature, recurrent neural network transducer (RNN-T) is a very promising end-to-end (E2E) model that may replace the popular hybrid model for automatic speech recognition.In this paper, we describe our recent development of RNN-T models with reduced GPU memory consumption during training, better initialization strategy, and advanced encoder modeling with future lookahead.When trained with Microsoft's 65 thousand hours of anonymized training data, the developed RNN-T model surpasses a very well trained hybrid model with both better recognition accuracy and lower latency.We further study how to customize RNN-T models to a new domain, which is important for deploying E2E models to practical scenarios.By comparing several methods leveraging text-only data in the new domain, we found that updating RNN-T's prediction and joint networks using text-to-speech generated from domain-specific text is the most effective.

While recurrent neural networks still largely define state-of-theart speech recognition systems, the Transformer network has been proven to be a competitive alternative, especially in the offline condition.Most studies with Transformers have been constrained in a relatively small scale setting, and some forms of data argumentation approaches are usually applied to combat the data sparsity issue.In this paper, we aim at understanding the behaviors of Transformers in the large-scale speech recognition setting, where we have used around 65,000 hours of training data.We investigated various aspects on scaling up Transformers, including model initialization, warmup training as well as different Layer Normalization strategies.In the streaming condition, we compared the widely used attention mask based future context lookahead approach to the Transformer-XL network.From our experiments, we show that Transformers can achieve around 6% relative word error rate (WER) reduction compared to the BLSTM baseline in the offline fashion, while in the streaming fashion, Transformer-XL is comparable to LC-BLSTM with 800 millisecond latency constraint.

Recent studies suggest that it may now be possible to construct end-to-end Neural Network (NN) models that perform on-par with, or even outperform, hybrid models in speech recognition. These models differ in their designs, and as such, may exhibit diverse and complementary error patterns. A combination between the predictions of these models may therefore yield significant gains. This paper studies the feasibility of performing hypothesis-level combination between hybrid and end-to-end NN models. The end-to-end NN models often exhibit a bias in their posteriors toward short hypotheses, and this may adversely affect Minimum Bayes’ Risk (MBR) combination methods. MBR training and length normalisation can be used to reduce this bias. Models are trained on Microsoft’s 75 thousand hours of anonymised data and evaluated on test sets with 1.8 million words. The results show that significant gains can be obtained by combining the hypotheses of hybrid and end-to-end NN models together.

Although mask based adaptive beamforming technique benefits speech recognition in far-field, noisy and multi-talker scenarios, it depends on the long time context to estimate target and interference statistics, thus when applied in applications with low latency requirement, its performance usually drops drastically. In contrast, the fixed beamformers do not import time delay but usually have limited capability in acoustic cancellation of interfering source. In this work, we propose a novel multi-channel speech separation system that targets at overlapped speech recognition with low latency processing, which includes four jointly optimized components: a pre-separator, a set of fixed beamformer, an attentional selection module and neural post filtering. With proposed model, low latency processing is achieved by utilizing the known microphone geometry information, while keeps the high quality separation through neural post filtering and end-to-end optimization. In our experiments, we show that the proposed system achieves comparable performance in offline evaluation with the mask based MVDR and speech extraction system, while yield remarkable improvements in the online evaluation.

In the last few years, an emerging trend in automatic speech recognition research is the study of end-to-end (E2E) systems. Connectionist Temporal Classification (CTC), Attention Encoder-Decoder (AED), and RNN Transducer (RNN-T) are the most popular three methods. Among these three methods, RNN-T has the advantages to do online streaming which is challenging to AED and it doesn't have CTC's frame-independence assumption. In this paper, we improve the RNN-T training in two aspects. First, we optimize the training algorithm of RNN-T to reduce the memory consumption so that we can have larger training minibatch for faster training speed. Second, we propose better model structures so that we obtain RNN-T models with the very good accuracy but small footprint. Trained with 30 thousand hours anonymized and transcribed Microsoft production data, the best RNN-T model with even smaller model size (216 Megabytes) achieves up-to 11.8% relative word error rate (WER) reduction from the baseline RNN-T model. This best RNN-T model is significantly better than the device hybrid model with similar size by achieving up-to 15.0% relative WER reduction, and obtains similar WERs as the server hybrid model of 5120 Megabytes in size.

Predicting words and subword units (WSUs) as the output has shown to be effective for the attention-based encoder-decoder (AED) model in end-to-end speech recognition. However, as one input to the decoder recurrent neural network (RNN), each WSU embedding is learned independently through context and acoustic information in a purely data-driven fashion. Little effort has been made to explicitly model the morphological relationships among WSUs. In this work, we propose a novel character-aware (CA) AED model in which each WSU embedding is computed by summarizing the embeddings of its constituent characters using a CA-RNN. This WSU-independent CA-RNN is jointly trained with the encoder, the decoder and the attention network of a conventional AED to predict WSUs. With CA-AED, the embeddings of morphologically similar WSUs are naturally and directly correlated through the CA-RNN in addition to the semantic and acoustic relations modeled by a traditional AED. Moreover, CA-AED significantly reduces the model parameters in a traditional AED by replacing the large pool of WSU embeddings with a much smaller set of character embeddings. On a 3400 hours Microsoft Cortana dataset, CA-AED achieves up to 11.9% relative WER improvement over a strong AED baseline with 27.1% fewer model parameters.

Teacher-student (T/S) has shown to be effective for domain adaptation of deep neural network acoustic models in hybrid speech recognition systems. In this work, we extend the T/S learning to large-scale unsupervised domain adaptation of an attention-based end-to-end (E2E) model through two levels of knowledge transfer: teacher's token posteriors as soft labels and one-best predictions as decoder guidance. To further improve T/S learning with the help of ground-truth labels, we propose adaptive T/S (AT/S) learning. Instead of conditionally choosing from either the teacher's soft token posteriors or the one-hot ground-truth label, in AT/S, the student always learns from both the teacher and the ground truth with a pair of adaptive weights assigned to the soft and one-hot labels quantifying the confidence on each of the knowledge sources. The confidence scores are dynamically estimated at each decoder step as a function of the soft and one-hot labels. With 3400 hours parallel close-talk and far-field Microsoft Cortana data for domain adaptation, T/S and AT/S achieves 6.3% and 10.3% relative word error rate improvement over a strong E2E model trained with the same amount of far-field data.

Overlapped speech is one of the main challenges in conversational speech applications such as meeting transcription. Blind speech separation and speech extraction are two common approaches to this problem. Both of them, however, suffer from limitations resulting from the lack of abilities to either leverage additional information or process multiple speakers simultaneously. In this work, we propose a novel method called speech separation using speaker inventory (SSUSI), which combines the advantages of both approaches and thus solves their problems. SSUSI makes use of a speaker inventory, i.e. a pool of pre-enrolled speaker signals, and jointly separates all participating speakers. This is achieved by a specially designed attention mechanism, eliminating the need for accurate speaker identities. Experimental results show that SSUSI outperforms permutation invariant training based blind speech separation by up to 48% relatively in word error rate (WER). Compared with speech extraction, SSUSI reduces computation time by up to 70% and improves the WER by more than 13% relatively.

Text-independent speaker verification imposes no constraints on the spoken content and usually needs long observations to make reliable prediction. In this paper, we propose two speaker embedding approaches by integrating the phonetic information into the attention-based residual convolutional neural network (CNN). Phonetic features are extracted from the bottleneck layer of a pretrained acoustic model. In implicit phonetic attention (IPA), the phonetic features are projected by a transformation network into multi-channel feature maps, and then combined with the raw acoustic features as the input of the CNN network. In explicit phonetic attention (EPA), the phonetic features are directly connected to the attentive pooling layer through a separate 1-dim CNN to generate the attention weights. With the incorporation of spoken content and attention mechanism, the system can not only distill the speaker-discriminant frames but also actively normalize the phonetic variations. Multi-head attention and discriminative objectives are further studied to improve the system. Experiments on the VoxCeleb corpus show our proposed system could outperform the state-of-the-art by around 43% relative.

Acoustic models are domain dependent and do not perform well if there is a mismatch between training and test conditions. As an alternative, the Mixture of Experts (MoE) model was introduced for multi-domain modeling. It combines the outputs of several domain specific models (or experts) using a gating network. However, one drawback is that the gating network directly uses raw features and is unaware of the state of the experts. In this work, we propose several alternatives to improve the MoE model. First, to make our MoE model state-aware, we use outputs of experts as inputs to the gating network. Then we show that vector based interpolation of the mixture weights is more effective than scalar interpolation. Second, we show that directly learning the mixture weights without using any complex gating is still effective. Finally, we introduce a hybrid attention model that uses the logits and mixture weights from the previous time step to generate the mixture weights at the current time. Our best proposed model outperforms a baseline model using LSTM based gating achieving about 20.48% relative reduction in word error rate (WER). Moreover, it beats an oracle model which picks the best expert for a given test condition.

In our recent work, we proposed a layer trajectory long short-term memory (ltLSTM) model which decouples the tasks of temporal modeling and senone classification with time-LSTMs and depth-LSTMs. The ltLSTM model achieved significant accuracy improvement over the traditional multi-layer LSTM models from our previous study. Considering the future context frames carrying valuable information for predicting the target label evidenced by the success of bi-directional LSTMs, in this work we investigate how to incorporate this kind of information with hidden vectors from either time-LSTM or depth-LSTM. Trained with 30 thousand hours of EN-US Microsoft internal data, the best ltLSTM model with future context frames can improve the baseline ltLSTM with up to 11.5% relative word error rate (WER) reduction and improve the baseline LSTM with up to 24.6% relative WER reduction across different tasks.

Although great progress has been made on end-to-end (E2E) models for monolingual and multilingual automatic speech recognition (ASR), there is no successful study for E2E models on the challenging intra-sentential code-switching (CS) ASR task to our best knowledge. In this paper, we propose an approach for CS ASR using E2E connectionist temporal classification (CTC) models. We use a frame-level language identification model to linearly adjust the posteriors of an E2E CTC model. We evaluate the proposed method on Microsoft live Chinese Cortana data with 7000 hours Chinese and English monolingual data and 300 hours CS data as the training data. Trained with only monolingual data without observing any CS data, the proposed method can obtain up to 6.3% relative word error rate (WER) reduction. In the scenario of training with both monolingual and CS data, the proposed method can get up to 4.2% relative WER improvement. This approach can also maintain comparable performance on a Chinese test set compared with baseline models.

We propose a novel adversarial speaker adaptation (ASA) scheme, in which adversarial learning is applied to regularize the distribution of deep hidden features in a speaker-dependent (SD) deep neural network (DNN) acoustic model to be close to that of a fixed speaker-independent (SI) DNN acoustic model during adaptation. An additional discriminator network is introduced to distinguish the deep features generated by the SD model from those produced by the SI model. In ASA, with a fixed SI model as the reference, an SD model is jointly optimized with the discriminator network to minimize the senone classification loss, and simultaneously to mini-maximize the SI/SD discrimination loss on the adaptation data. With ASA, a senone-discriminative deep feature is learned in the SD model with a similar distribution to that of the SI model. With such a regularized and adapted deep feature, the SD model can perform improved automatic speech recognition on the target speaker's speech. Evaluated on the Microsoft short message dictation dataset, ASA achieves 14.4% and 7.9% relative word error rate improvements for supervised and unsupervised adaptation, respectively, over an SI model trained from 2600 hours data, with 200 adaptation utterances per speaker.

Adversarial domain-invariant training (ADIT) proves to be effective in suppressing the effects of domain variability in acoustic modeling and has led to improved performance in automatic speech recognition (ASR). In ADIT, an auxiliary domain classifier takes in equally-weighted deep features from a deep neural network (DNN) acoustic model and is trained to improve their domain-invariance by optimizing an adversarial loss function. In this work, we propose an attentive ADIT (AADIT) in which we advance the domain classifier with an attention mechanism to automatically weight the input deep features according to their importance in domain classification. With this attentive re-weighting, ADDIT can focus on the domain normalization of phonetic components that are more susceptible to domain variability and generates deep features with improved domain-invariance and senone-discriminativity over ADIT. Most importantly, the attention block serves only as an external component to the DNN acoustic model and is not involved in ASR, so AADIT can be used to improve the acoustic modeling with any DNN architectures. More generally, the same methodology can improve any adversarial learning system with an auxiliary discriminator. Evaluated on CHiME-3 dataset, the AADIT achieves 13.6% and 9.3% relative WER improvements, respectively, over a multi-conditional model and a strong ADIT baseline.

The teacher-student (T/S) learning has been shown to be effective for a variety of problems such as domain adaptation and model compression. One shortcoming of the T/S learning is that a teacher model, not always perfect, sporadically produces wrong guidance in form of posterior probabilities that misleads the student model towards a suboptimal performance. To overcome this problem, we propose a conditional T/S learning scheme, in which a "smart" student model selectively chooses to learn from either the teacher model or the ground truth labels conditioned on whether the teacher can correctly predict the ground truth. Unlike a naive linear combination of the two knowledge sources, the conditional learning is exclusively engaged with the teacher model when the teacher model's prediction is correct, and otherwise backs off to the ground truth. Thus, the student model is able to learn effectively from the teacher and even potentially surpass the teacher. We examine the proposed learning scheme on two tasks: domain adaptation on CHiME-3 dataset and speaker adaptation on Microsoft short message dictation dataset. The proposed method achieves 9.8% and 12.8% relative word error rate reductions, respectively, over T/S learning for environment adaptation and speaker-independent model for speaker adaptation.

The use of deep networks to extract embeddings for speaker recognition has proven successfully. However, such embeddings are susceptible to performance degradation due to the mismatches among the training, enrollment, and test conditions. In this work, we propose an adversarial speaker verification (ASV) scheme to learn the condition-invariant deep embedding via adversarial multi-task training. In ASV, a speaker classification network and a condition identification network are jointly optimized to minimize the speaker classification loss and simultaneously mini-maximize the condition loss. The target labels of the condition network can be categorical (environment types) and continuous (SNR values). We further propose multi-factorial ASV to simultaneously suppress multiple factors that constitute the condition variability. Evaluated on a Microsoft Cortana text-dependent speaker verification task, the ASV achieves 8.8% and 14.5% relative improvements in equal error rates (EER) for known and unknown conditions, respectively.

Directly emitting words and sub-words from speech spectrogram has been shown to produce good results using end-to-end (E2E) trained models. Connectionist Temporal Classification (CTC) and Sequence-to-Sequence attention (Seq2Seq) models have both shown better success when directly targeting words or sub-words. In this work, we ask the question: Can an E2E model go beyond words and transcribe directly to phrases (i.e., a group of words)? Directly modeling frequent phrases might be better than modeling its constituent words. Also, emitting multiple words together might speed up inference in models like Seq2Seq where decoding is inherently sequential. To answer this, we undertake a study on a 3400-hour Microsoft Cortana voice assistant task. We present a side-by-side comparison for CTC and Seq2Seq models that have been trained to target a variety of tokens including letters, sub-words, words and phrases. We show that an E2E model can indeed transcribe directly to phrases. We see that while CTC has difficulty in accurately modeling phrases, a more powerful model like Seq2Seq can effortlessly target phrases that are up to 4 words long, with only a reasonable degradation in the final word error rate.

We propose three regularization-based speaker adaptation approaches to adapt the attention-based encoder-decoder (AED) model with very limited adaptation data from target speakers for end-to-end automatic speech recognition. The first method is Kullback-Leibler divergence (KLD) regularization, in which the output distribution of a speaker-dependent (SD) AED is forced to be close to that of the speaker-independent (SI) model by adding a KLD regularization to the adaptation criterion. To compensate for the asymmetric deficiency in KLD regularization, an adversarial speaker adaptation (ASA) method is proposed to regularize the deep-feature distribution of the SD AED through the adversarial learning of an auxiliary discriminator and the SD AED. The third approach is the multi-task learning, in which an SD AED is trained to jointly perform the primary task of predicting a large number of output units and an auxiliary task of predicting a small number of output units to alleviate the target sparsity issue. Evaluated on a Microsoft short message dictation task, all three methods are highly effective in adapting the AED model, achieving up to 12.2% and 3.0% word error rate improvement over an SI AED trained from 3400 hours data for supervised and unsupervised adaptation, respectively.

The acoustic-to-word model based on the Connectionist Temporal Classification (CTC) criterion is a natural end-to-end (E2E) system directly targeting word as output unit. Two issues exist in the system: first, the current output of the CTC model relies on the current input and does not account for context weighted inputs. This is the hard alignment issue. Second, the word-based CTC model suffers from the out-of-vocabulary (OOV) issue. This means it can model only frequently occurring words while tagging the remaining words as OOV. Hence, such a model is limited in its capacity in recognizing only a fixed set of frequent words. In this study, we propose addressing these problems using a combination of attention mechanism and mixed-units. In particular, we introduce Attention CTC, Self-Attention CTC, Hybrid CTC, and Mixed-unit CTC. First, we blend attention modeling capabilities directly into the CTC network using Attention CTC and Self-Attention CTC. Second, to alleviate the OOV issue, we present Hybrid CTC which uses a word and letter CTC with shared hidden layers. The Hybrid CTC consults the letter CTC when the word CTC emits an OOV. Then, we propose a much better solution by training a Mixed-unit CTC which decomposes all the OOV words into sequences of frequent words and multi-letter units. Evaluated on a 3400 hours Microsoft Cortana voice assistant task, our final acoustic-to-word solution using attention and mixed-units achieves a relative reduction in word error rate (WER) over the vanilla word CTC by 12.09%. Such an E2E model without using any language model (LM) or complex decoder also outperforms a traditional context-dependent (CD) phoneme CTC with strong LM and decoder by 6.79% relative.

In this study, we propose advancing all-neural speech recognition by directly incorporating attention modeling within the Connectionist Temporal Classification (CTC) framework. In particular, we derive new context vectors using time convolution features to model attention as part of the CTC network. To further improve attention modeling, we utilize content information extracted from a network representing an implicit language model. Finally, we introduce vector based attention weights that are applied on context vectors across both time and their individual components. We evaluate our system on a 3400 hours Microsoft Cortana voice assistant task and demonstrate that our proposed model consistently outperforms the baseline model achieving about 20% relative reduction in word error rates.

The acoustic-to-word model based on the connectionist temporal classification (CTC) criterion was shown as a natural end-to-end (E2E) model directly targeting words as output units. However, the word-based CTC model suffers from the out-of-vocabulary (OOV) issue as it can only model limited number of words in the output layer and maps all the remaining words into an OOV output node. Hence, such a word-based CTC model can only recognize the frequent words modeled by the network output nodes. Our first attempt to improve the acoustic-to-word model is a hybrid CTC model which consults a letter-based CTC when the word-based CTC model emits OOV tokens during testing time. Then, we propose a much better solution by training a mixed-unit CTC model which decomposes all the OOV words into sequences of frequent words and multi-letter units. Evaluated on a 3400 hours Microsoft Cortana voice assistant task, the final acoustic-to-word solution improves the baseline word-based CTC by relative 12.09% word error rate (WER) reduction when combined with our proposed attention CTC. Such an E2E model without using any language model (LM) or complex decoder outperforms the traditional context-dependent phoneme CTC which has strong LM and decoder by relative 6.79%.

In this study, we develop the keyword spotting (KWS) and acoustic model (AM) components in a far-field speaker system. Specifically, we use teacher-student (T/S) learning to adapt a close-talk well-trained production AM to far-field by using parallel close-talk and simulated far-field data. We also use T/S learning to compress a large-size KWS model into a small-size one to fit the device computational cost. Without the need of transcription, T/S learning well utilizes untranscribed data to boost the model performance in both the AM adaptation and KWS model compression. We further optimize the models with sequence discriminative training and live data to reach the best performance of systems. The adapted AM improved from the baseline by 72.60% and 57.16% relative word error rate reduction on play-back and live test data, respectively. The final KWS model size was reduced by 27 times from a large-size KWS model without losing accuracy.

We propose a novel adversarial multi-task learning scheme, aiming at actively curtailing the inter-talker feature variability while maximizing its senone discriminability so as to enhance the performance of a deep neural network (DNN) based ASR system. We call the scheme speaker-invariant training (SIT). In SIT, a DNN acoustic model and a speaker classifier network are jointly optimized to minimize the senone (tied triphone state) classification loss, and simultaneously mini-maximize the speaker classification loss. A speaker-invariant and senone-discriminative deep feature is learned through this adversarial multi-task learning. With SIT, a canonical DNN acoustic model with significantly reduced variance in its output probabilities is learned with no explicit speaker-independent (SI) transformations or speaker-specific representations used in training or testing. Evaluated on the CHiME-3 dataset, the SIT achieves 4.99% relative word error rate (WER) improvement over the conventional SI acoustic model. With additional unsupervised speaker adaptation, the speaker-adapted (SA) SIT model achieves 4.86% relative WER gain over the SA SI acoustic model.

The teacher-student (T/S) learning has been shown effective in unsupervised domain adaptation [1]. It is a form of transfer learning, not in terms of the transfer of recognition decisions, but the knowledge of posteriori probabilities in the source domain as evaluated by the teacher model. It learns to handle the speaker and environment variability inherent in and restricted to the speech signal in the target domain without proactively addressing the robustness to other likely conditions. Performance degradation may thus ensue. In this work, we advance T/S learning by proposing adversarial T/S learning to explicitly achieve condition-robust unsupervised domain adaptation. In this method, a student acoustic model and a condition classifier are jointly optimized to minimize the Kullback-Leibler divergence between the output distributions of the teacher and student models, and simultaneously, to min-maximize the condition classification loss. A condition-invariant deep feature is learned in the adapted student model through this procedure. We further propose multi-factorial adversarial T/S learning which suppresses condition variabilities caused by multiple factors simultaneously. Evaluated with the noisy CHiME-3 test set, the proposed methods achieve relative word error rate improvements of 44.60% and 5.38%, respectively, over a clean source model and a strong T/S learning baseline model.

Voice assistant represents one of the most popular and important scenarios for speech recognition. In this paper, we propose two adaptation approaches to customize a multi-style well-trained acoustic model towards its subsidiary domain of Cortana assistant. First, we present anchor-based speaker adaptation by extracting the speaker information, i-vector or d-vector embeddings, from the anchor segments of ‘Hey Cortana’. The anchor embeddings are mapped to layer-wise parameters to control the transformations of both weight matrices and biases of multiple layers. Second, we directly update the existing model parameters for domain adaptation. We demonstrate that prior distribution should be updated along with the network adaptation to compensate the label bias from the development data. Updating the priors may have a significant impact when the target domain features high occurrence of anchor words. Experiments on Hey Cortana desktop test set show that both approaches improve the recognition accuracy significantly. The anchor-based adaptation using the anchor d-vector and the prior interpolation achieves 32% relative reduction in WER over the generic model.

Achieving high accuracy with end-to-end speech recognizers requires careful parameter initialization prior to training.Otherwise, the networks may fail to find a good local optimum.This is particularly true for online networks, such as unidirectional LSTMs.Currently, the best strategy to train such systems is to bootstrap the training from a tied-triphone system.However, this is time consuming, and more importantly, is impossible for languages without a high-quality pronunciation lexicon.In this work, we propose an initialization strategy that uses teacher-student learning to transfer knowledge from a large, well-trained, offline end-to-end speech recognition model to an online end-to-end model, eliminating the need for a lexicon or any other linguistic resources.We also explore curriculum learning and label smoothing and show how they can be combined with the proposed teacher-student learning for further improvements.We evaluate our methods on a Microsoft Cortana personal assistant task and show that the proposed method results in a 19% relative improvement in word error rate compared to a randomly-initialized baseline system.

It is popular to stack LSTM layers to get better modeling power, especially when large amount of training data is available. However, an LSTM-RNN with too many vanilla LSTM layers is very hard to train and there still exists the gradient vanishing issue if the network goes too deep. This issue can be partially solved by adding skip connections between layers, such as residual LSTM. In this paper, we propose a layer trajectory LSTM (ltLSTM) which builds a layer-LSTM using all the layer outputs from a standard multi-layer time-LSTM. This layer-LSTM scans the outputs from time-LSTMs, and uses the summarized layer trajectory information for final senone classification. The forward-propagation of time-LSTM and layer-LSTM can be handled in two separate threads in parallel so that the network computation time is the same as the standard time-LSTM. With a layer-LSTM running through layers, a gated path is provided from the output layer to the bottom layer, alleviating the gradient vanishing issue. Trained with 30 thousand hours of EN-US Microsoft internal data, the proposed ltLSTM performed significantly better than the standard multi-layer LSTM and residual LSTM, with up to 9.0% relative word error rate reduction across different tasks.

Feature mapping using deep neural networks is an effective approach for single-channel speech enhancement. Noisy features are transformed to the enhanced ones through a mapping network and the mean square errors between the enhanced and clean features are minimized. In this paper, we propose a cycle-consistent speech enhancement (CSE) in which an additional inverse mapping network is introduced to reconstruct the noisy features from the enhanced ones. A cycle-consistent constraint is enforced to minimize the reconstruction loss. Similarly, a backward cycle of mappings is performed in the opposite direction with the same networks and losses. With cycle-consistency, the speech structure is well preserved in the enhanced features while noise is effectively reduced such that the feature-mapping network generalizes better to unseen data. In cases where only unparalleled noisy and clean data is available for training, two discriminator networks are used to distinguish the enhanced and noised features from the clean and noisy ones. The discrimination losses are jointly optimized with reconstruction losses through adversarial multi-task learning. Evaluated on the CHiME-3 dataset, the proposed CSE achieves 19.60% and 6.69% relative word error rate improvements respectively when using or without using parallel clean and noisy speech data.

Feature-mapping with deep neural networks is commonly used for single-channel speech enhancement, in which a feature-mapping network directly transforms the noisy features to the corresponding enhanced ones and is trained to minimize the mean square errors between the enhanced and clean features. In this paper, we propose an adversarial feature-mapping (AFM) method for speech enhancement which advances the feature-mapping approach with adversarial learning. An additional discriminator network is introduced to distinguish the enhanced features from the real clean ones. The two networks are jointly optimized to minimize the feature-mapping loss and simultaneously mini-maximize the discrimination loss. The distribution of the enhanced features is further pushed towards that of the clean features through this adversarial multi-task training. To achieve better performance on ASR task, senone-aware (SA) AFM is further proposed in which an acoustic model network is jointly trained with the feature-mapping and discriminator networks to optimize the senone classification loss in addition to the AFM losses. Evaluated on the CHiME-3 dataset, the proposed AFM achieves 16.95% and 5.27% relative word error rate (WER) improvements over the real noisy data and the feature-mapping baseline respectively and the SA-AFM achieves 9.85% relative WER improvement over the multi-conditional acoustic model.

Although advances in close-talk speech recognition have resulted in relatively low error rates, the recognition performance in far-field environments is still limited due to low signal-to-noise ratio, reverberation, and overlapped speech from simultaneous speakers which is especially more difficult. To solve these problems, beamforming and speech separation networks were previously proposed. However, they tend to suffer from leakage of interfering speech or limited generalizability. In this work, we propose a simple yet effective method for multi-channel far-field overlapped speech recognition. In the proposed system, three different features are formed for each target speaker, namely, spectral, spatial, and angle features. Then a neural network is trained using all features with a target of the clean speech of the required speaker. An iterative update procedure is proposed in which the mask-based beamforming and mask estimation are performed alternatively. The proposed system were evaluated with real recorded meetings with different levels of overlapping ratios. The results show that the proposed system achieves more than 24% relative word error rate (WER) reduction than fixed beamforming with oracle selection. Moreover, as overlap ratio rises from 20% to 70+%, only 3.8% WER increase is observed for the proposed system.

Traditional LSTM model and its variants normally work in a frame-by-frame and layer-by-layer fashion, which deals with the temporal modeling and target classification problems at the same time. In this paper, we extend our recently proposed layer trajectory LSTM (ltLSTM) and present a generalized framework, which is equipped with a depth processing block that scans the hidden states of each time-LSTM layer, and uses the summarized layer trajectory information for final senone classification. We explore different modeling units used in the depth processing block to have a good tradeoff between accuracy and runtime cost. Furthermore, we integrate an attention module into this framework to explore wide context information, which is especially beneficial for uni-directional LSTMs. Trained with 30 thousand hours of EN-US Microsoft internal data and cross entropy criterion, the proposed generalized ltLSTM performed significantly better than the standard multi-layer time-LSTM, with up to 12.8% relative word error rate (WER) reduction across different tasks. With attention modeling, the relative WER reduction can be up to 17.9%. We observed similar gain when the models were trained with sequence discriminative training criterion.

We propose two approaches for speaker adaptation in end-to-end (E2E) automatic speech recognition systems. One is Kullback-Leibler divergence (KLD) regularization and the other is multi-task learning (MTL). Both approaches aim to address the data sparsity especially output target sparsity issue of speaker adaptation in E2E systems. The KLD regularization adapts a model by forcing the output distribution from the adapted model to be close to the unadapted one. The MTL utilizes a jointly trained auxiliary task to improve the performance of the main task. We investigated our approaches on E2E connectionist temporal classification (CTC) models with three different types of output units. Experiments on the Microsoft short message dictation task demonstrated that MTL outperforms KLD regularization. In particular, the MTL adaptation obtained 8.8% and 4.0% relative word error rate reductions (WERRs) for supervised and unsupervised adaptations for the word CTC model, and 9.6% and 3.8% relative WERRs for the mix-unit CTC model, respectively.

Unsupervised single-channel overlapped speech recognition is one of the hardest problems in automatic speech recognition (ASR). Permutation invariant training (PIT) is a state of the art model-based approach, which applies a single neural network to solve this single-input, multiple-output modeling problem. We propose to advance the current state of the art by imposing a modular structure on the neural network, applying a progressive pretraining regimen, and improving the objective function with transfer learning and a discriminative training criterion. The modular structure splits the problem into three subtasks: frame-wise interpreting, utterance-level speaker tracing, and speech recognition. The pretraining regimen uses these modules to solve progressively harder tasks. Transfer learning leverages parallel clean speech to improve the training targets for the network. Our discriminative training formulation is a modification of standard formulations that also penalizes competing outputs of the system. Experiments are conducted on the artificial overlapped switchboard and hub5e-swb dataset. The proposed framework achieves over 30% relative improvement of word error rate over both a strong jointly trained system, PIT for ASR, and a separately optimized system, PIT for speech separation with clean speech ASR model. The improvement comes from better model generalization, training efficiency, and the sequence level linguistic knowledge integration.

While recent progresses in neural network approaches to singlechannel speech separation, or more generally the cocktail party problem, achieved significant improvement, their performance for complex mixtures is still not satisfactory. In this work, we propose a novel multi-channel framework for multi-talker separation. In the proposed model, an input multi-channel mixture signal is firstly converted to a set of beamformed signals using fixed beam patterns. For this beamforming, we propose to use differential beamformers as they are more suitable for speech separation. Then each beamformed signal is fed into a single-channel anchored deep attractor network to generate separated signals. And the final separation is acquired by post selecting the separating output for each beams. To evaluate the proposed system, we create a challenging dataset comprising mixtures of 2, 3 or 4 speakers. Our results show that the proposed system largely improves the state of the art in speech separation, achieving 11.5 dB, 11.76 dB and 11.02 dB average signal-to-distortion ratio improvement for 4, 3 and 2 overlapped speaker mixtures, which is comparable to the performance of a minimum variance distortionless response beamformer that uses oracle location, source, and noise information. We also run speech recognition with a clean trained acoustic model on the separated speech, achieving relative word error rate (WER) reduction of 45.76%, 59.40% and 62.80% on fully overlapped speech of 4, 3 and 2 speakers, respectively. With a far talk acoustic model, the WER is further reduced.

Recently, the acoustic-to-word model based on the Connectionist Temporal Classification (CTC) criterion was shown as a natural end-to-end model directly targeting words as output units. However, this type of word-based CTC model suffers from the out-of-vocabulary (OOV) issue as it can only model limited number of words in the output layer and maps all the remaining words into an OOV output node. Therefore, such word-based CTC model can only recognize the frequent words modeled by the network output nodes. It also cannot easily handle the hot-words which emerge after the model is trained. In this study, we improve the acoustic-to-word model with a hybrid CTC model which can predict both words and characters at the same time. With a shared-hidden-layer structure and modular design, the alignments of words generated from the word-based CTC and the character-based CTC are synchronized. Whenever the acoustic-to-word model emits an OOV token, we back off that OOV segment to the word output generated from the character-based CTC, hence solving the OOV or hot-words issue. Evaluated on a Microsoft Cortana voice assistant task, the proposed model can reduce the errors introduced by the OOV output token in the acoustic-to-word model by 30%.

Unsupervised domain adaptation of speech signal aims at adapting a well-trained source-domain acoustic model to the unlabeled data from target domain. This can be achieved by adversarial training of deep neural network (DNN) acoustic models to learn an intermediate deep representation that is both senone-discriminative and domain-invariant. Specifically, the DNN is trained to jointly optimize the primary task of senone classification and the secondary task of domain classification with adversarial objective functions. In this work, instead of only focusing on learning a domain-invariant feature (i.e. the shared component between domains), we also characterize the difference between the source and target domain distributions by explicitly modeling the private component of each domain through a private component extractor DNN. The private component is trained to be orthogonal with the shared component and thus implicitly increases the degree of domain-invariance of the shared component. A reconstructor DNN is used to reconstruct the original speech feature from the private and shared components as a regularization. This domain separation framework is applied to the unsupervised environment adaptation task and achieved 11.08% relative WER reduction from the gradient reversal layer training, a representative adversarial training method, for automatic speech recognition on CHiME-3 dataset.

In the era of deep learning, although beam-forming multi-channel signal processing is still very helpful, it was reported that single-channel robust front-ends usually cannot benefit deep learning models because the layer-by-layer structure of deep learning models provides a feature extraction strategy that automatically derives powerful noise-resistant features from primitive raw data for senone classification. In this study, we show that the single-channel robust front-end is still very beneficial to deep learning modelling as long as it is well designed. We improve a robust front-end, cepstra minimum mean square error (CMMSE), by using more reliable voice activity detector, refined prior SNR estimation, better gain smoothing and two-stage processing. This new front-end, improved CMMSE (ICMMSE), is evaluated on the standard Aurora 2 and Chime 3 tasks, and a 3400 hour Microsoft Cortana digital assistant task using Gaussian mixture models, feed-forward deep neural networks, and long short-term memory recurrent neural networks, respectively. It is shown that ICMMSE is superior regardless of the underlying acoustic models and the scale of evaluation tasks, with 25.46% relative WER reduction on Aurora 2, up to 11.98% relative WER reduction on Chime 3, and up to 11.01% relative WER reduction on Cortana digital assistant task, respectively.

Recently, the low-rank plus diagonal (LRPD) adaptation was proposed for speaker adaptation of deep neural network (DNN) models. The LRPD restructures the adaptation matrix as a superposition of a diagonal matrix and a product of two low-rank matrices. In this paper, we extend the LRPD adaptation into the subspace-based approach to further reduce the speaker-dependent (SD) footprint. We apply the extended LRPD (eLRPD) adaptation for the DNN and LSTM models with emphasis placed on the applicability of the adaptation to large-scale speech recognition systems. To speed up the adaptation in test time, we propose the bottleneck (BN) caching approach to eliminate the redundant computations during multiple sweeps of development data. Experimental results on the short message dictation (SMD) task show that the eLRPD adaptation can reduce the SD footprints by 82% for the SVD DNN and 96% for the LSTM-RNN over the linear adaptation, while maintaining the comparable accuracy. The BN caching achieves up to 3.5 times speedup in adaptation at no loss of recognition accuracy.

For single-channel speech enhancement, mask learning based approach through neural network has been shown to outperform the feature mapping approach, and to be effective as a pre-processor for automatic speech recognition. However, its assumption that the mixture and clean reference must have the correspondent scale doesn’t hold in data collected from real world, and thus leads to significant performance degradation on parallel recorded data. In this paper, we first extend the mask learning based speech enhancement by integrating two types of restoration layer to address the scale mismatch problem. We further propose a novel residual learning based speech enhancement model via adding different shortcut connections to a feature mapping network. We show such a structure can benefit from both the mask learning and the feature mapping. We evaluate the proposed speech enhancement models on CHiME 3 data. Without retraining the acoustic model, the best bidirection LSTM with residue connections yields 24.90% relative WER reduction on real data and 34.57% WER on simulated data.

High accuracy speech recognition requires a large amount of transcribed data for supervised training.In the absence of such data, domain adaptation of a well-trained acoustic model can be performed, but even here, high accuracy usually requires significant labeled data from the target domain.In this work, we propose an approach to domain adaptation that does not require transcriptions but instead uses a corpus of unlabeled parallel data, consisting of pairs of samples from the source domain of the well-trained model and the desired target domain.To perform adaptation, we employ teacher/student (T/S) learning, in which the posterior probabilities generated by the source-domain model can be used in lieu of labels to train the target-domain model.We evaluate the proposed approach in two scenarios, adapting a clean acoustic model to noisy speech and adapting an adults' speech acoustic model to children's speech.Significant improvements in accuracy are obtained, with reductions in word error rate of up to 44% over the original source model without the need for transcribed data in the target domain.Moreover, we show that increasing the amount of unlabeled data results in additional model robustness, which is particularly beneficial when using simulated training data in the target-domain.

Long short-term memory (LSTM) recurrent neural networks (RNNs) have recently shown significant performance improvements over deep feed-forward neural networks. A key aspect of these models is the use of time recurrence, combined with a gating architecture that allows them to track the long-term dynamics of speech. Inspired by human spectrogram reading, we recently proposed the frequency LSTM (F-LSTM) that performs 1-D recurrence over the frequency axis and then performs 1-D recurrence over the time axis. In this study, we further improve the acoustic model by proposing a 2-D, time-frequency (TF) LSTM. The TF-LSTM jointly scans the input over the time and frequency axes to model spectro-temporal warping, and then uses the output activations as the input to a time LSTM (T-LSTM). The joint time-frequency modeling better normalizes the features for the upper layer T-LSTMs. Evaluated on a 375-hour short message dictation task, the proposed TF-LSTM obtained a 3.4% relative WER reduction over the best T-LSTM. The invariance property achieved by joint time-frequency analysis is demonstrated on a mismatched test set, where the TF-LSTM achieves a 14.2% relative WER reduction over the best T-LSTM.

On acoustic modeling, recurrent neural networks (RNNs) using Long Short-Term Memory (LSTM) units have recently been shown to outperform deep neural networks (DNNs) models. This paper focuses on resolving two challenges faced by LSTM models: high model complexity and poor decoding efficiency. Motivated by our analysis of the gates activation and function, we present two LSTM simplifications: deriving input gates from forget gates, and removing recurrent inputs from output gates. To accelerate decoding of LSTMs, we propose to apply frame skipping during training, and frame skipping and posterior copying (FSPC) during decoding. In the experiments, model simplifications reduce the size of LSTM models by 26%, resulting in a simpler model structure. Meanwhile, the application of FSPC speeds up model computation by 2 times during LSTM decoding. All these improvements are achieved at the cost of 1% WER degradation.

Recurrent Neural Networks (RNNs) using Long-Short Term Memory (LSTM) architecture have demonstrated the state-of-the-art performances on speech recognition. Most of deep RNNs use the softmax activation function in the last layer for classification. This paper illustrates small but consistent advantages of replacing the softmax layer in RNN with Support Vector Machines (SVMs). The parameters of RNNs and SVMs are jointly learned using a sequence-level max-margin criteria, instead of cross-entropy. The resulting model is termed Recurrent SVM. The conventional SVMs need to predefine a feature space and do not have internal states to deal with arbitrary long-term dependencies in sequences. The proposed recurrent SVM uses LSTMs to learn the feature space and to capture temporal dependencies, while using the SVM (in the last layer) for sequence classification. The model is evaluated on the Windows phone task for large vocabulary continuous speech recognition.

In this paper, we propose a scalable adaptation technique that adapts the deep neural network (DNN) model through the low-rank plus diagonal (LRPD) decomposition. It is desired that an adaptation method can properly accommodate the available development data with a variable amount of adaptation parameters. Thus, the resulting models neither over-fit nor under-fit as the development data vary in size for different speakers. The technique developed in this paper is inspired by observing that adaptation matrices are very close to an identity matrix or diagonally dominant. The LRPD restructures the adaptation matrix as a superposition of a diagonal matrix and a low-rank matrix. By varying the low-rank values, the LRPD contains the full and the diagonal adaptation matrix as its special cases. Experimental results demonstrated that the LRPD adaptation of the full-size DNN obtains improved accuracy over the standard linear adaptation. The LRPD bottleneck adaptation can reduce the speaker-specific footprint by 82% over an already very compact SVD bottleneck adaptation, at an expense of 1% relative WER increase.

In this paper, we propose a deep convolutional neural network (CNN) with layer-wise context expansion and location-based attention, for large vocabulary speech recognition. In our model each higher layer uses information from broader contexts, along both the time and frequency dimensions, than its immediate lower layer. We show that both the layer-wise context expansion and the location-based attention can be implemented using the element-wise matrix product and the convolution operation. For this reason, contrary to other CNNs, no pooling operation is used in our model. Experiments on the 309hr Switchboard task and the 375hr short message dictation task indicates that our model outperforms both the DNN and LSTM significantly.

A new type of End-to-End system for text-dependent speaker verification is presented in this paper. Previously, using the phonetic discriminate/speaker discriminate DNN as a feature extractor for speaker verification has shown promising results. The extracted frame-level (bottleneck, posterior or d-vector) features are equally weighted and aggregated to compute an utterance-level speaker representation (d-vector or i-vector). In this work we use a speaker discriminate CNN to extract the noise-robust frame-level features. These features are smartly combined to form an utterance-level speaker vector through an attention mechanism. The proposed attention model takes the speaker discriminate information and the phonetic information to learn the weights. The whole system, including the CNN and attention model, is joint optimized using an end-to-end criterion. The training algorithm imitates exactly the evaluation process — directly mapping a test utterance and a few target speaker utterances into a single verification score. The algorithm can smartly select the most similar impostor for each target speaker to train the network. We demonstrated the effectiveness of the proposed end-to-end system on Windows 10 “Hey Cortana” speaker verification task.

This work presents a broad study on the adaptation of neural network acoustic models by means of learning hidden unit contributions (LHUC) - a method that linearly re-combines hidden units in a speaker- or environment-dependent manner using small amounts of unsupervised adaptation data. We also extend LHUC to a speaker adaptive training (SAT) framework that leads to a more adaptable DNN acoustic model, working both in a speaker-dependent and a speaker-independent manner, without the requirements to maintain auxiliary speaker-dependent feature extractors or to introduce significant speaker-dependent changes to the DNN structure. Through a series of experiments on four different speech recognition benchmarks (TED talks, Switchboard, AMI meetings, and Aurora4) comprising 270 test speakers, we show that LHUC in both its test-only and SAT variants results in consistent word error rate reductions ranging from 5% to 23% relative depending on the task and the degree of mismatch between training and test data. In addition, we have investigated the effect of the amount of adaptation data per speaker, the quality of unsupervised adaptation targets, the complementarity to other adaptation techniques, one-shot adaptation, and an extension to adapting DNNs trained in a sequence discriminative manner.

Long short-term memory (LSTM) recurrent neural networks (RNNs) have recently shown significant performance improvements over deep feed-forward neural networks (DNNs). A key aspect of these models is the use of time recurrence, combined with a gating architecture that ameliorates the vanishing gradient problem. Inspired by human spectrogram reading, in this paper we propose an extension to LSTMs that performs the recurrence in frequency as well as in time. This model first scans the frequency bands to generate a summary of the spectral information, and then uses the output layer activations as the input to a traditional time LSTM (T-LSTM). Evaluated on a Microsoft short message dictation task, the proposed model obtained a 3.6% relative word error rate reduction over the T-LSTM.

Despite the fact that several sites have reported the effectiveness of convolutional neural networks (CNNs) on some tasks, there is no deep analysis regarding why CNNs perform well and in which case we should see CNNs' advantage. In the light of this, this paper aims to provide some detailed analysis of CNNs. By visualizing the localized filters learned in the convolutional layer, we show that edge detectors in varying directions can be automatically learned. We then identify four domains we think CNNs can consistently provide advantages over fully-connected deep neural networks (DNNs): channel-mismatched training-test conditions, noise robustness, distant speech recognition, and low-footprint models. For distant speech recognition, a CNN trained on 1000 hours of Kinect distant speech data obtains relative 4% word error rate reduction (WERR) over a DNN of a similar size. To our knowledge, this is the largest corpus so far reported in the literature for CNNs to show its effectiveness. Lastly, we establish that the CNN structure combined with maxout units is the most effective model under small-sizing constraints for the purpose of deploying small-footprint models to devices. This setup gives relative 9.3% WERR from DNNs with sigmoid units.

Due to a large number of parameters in deep neural networks (DNNs), it is challenging to design a small-footprint DNN-based speech recognition system while maintaining a high recognition performance. Even with a singular value matrix decomposition (SVD) method and scalar quantization, the DNN model is still too large to be deployed on many mobile devices. Common practices like reducing the number of hidden nodes often result in significant accuracy loss. In this work, we propose to split each row vector of weight matrices into sub-vectors, and quantize them into a set of codewords using a split vector quantization (split-VQ) algorithm. The codebook can be fine-tuned using back-propagation when an aggressive quantization is performed. Experimental results demonstrate that the proposed method can further reduce the model size by 75% to 80% and save 10% to 50% computation on top of an already very compact SVD-DNN without a noticeable performance degradation. This results in a 3.2 MB-footprint DNN giving similar recognition performance as what a 59.1 MB standard DNN can achieve.

To develop speaker adaptation algorithms for deep neural network (DNN) that are suitable for large-scale online deployment, it is desirable that the adaptation model be represented in a compact form and learned in an unsupervised fashion. In this paper, we propose a novel low-footprint adaptation technique for DNN that adapts the DNN model through node activation functions. The approach introduces slope and bias parameters in the sigmoid activation functions for each speaker, allowing the adaptation model to be stored in a small-sized storage space. We show that this adaptation technique can be formulated in a linear regression fashion, analogous to other speak adaptation algorithms that apply additional linear transformations to the DNN layers. We further investigate semi-supervised online adaptation by making use of the user click-through data as a supervision signal. The proposed method is evaluated on short message dictation and voice search tasks in supervised, unsupervised, and semi-supervised setups. Compared with the singular value decomposition (SVD) bottleneck adaptation, the proposed adaptation method achieves comparable accuracy improvements with much smaller footprint.

We propose a novel approach to addressing the adaptation effectiveness issue in parameter adaptation for deep neural network (DNN) based acoustic models for automatic speech recognition by adding one or more small auxiliary output layers modeling broad acoustic units, such as mono-phones or tied-state (often called senone) clusters. In scenarios with a limited amount of available adaptation data, most senones are usually rarely seen or not observed, and consequently the ability to model them in a new condition is often not fully exploited. With the original senone classification task as the primary task, and adding auxiliary mono-phone/senone-cluster classification as the secondary tasks, multi-task learning (MTL) is employed to adapt the DNN parameters. With the proposed MTL adaptation framework, we improve the learning ability of the original DNN structure, then enlarge the coverage of the acoustic space to deal with the unseen senone problem, and thus enhance the discrimination power of the adapted DNN models. Experimental results on the 20,000-word open vocabulary WSJ task demonstrate that the proposed framework consistently outperforms the conventional linear hidden layer adaptation schemes without MTL by providing 5.4% relative reduction in word error rate (WERR) with only 1 single adaptation utterance, and 10.7% WERR with 40 adaptation utterances against the un-adapted DNN models.

We present a Bayesian approach to adapting parameters of a well-trained context-dependent, deep-neural-network, hidden Markov model (CD-DNN-HMM) to improve automatic speech recognition performance. Given an abundance of DNN parameters but with only a limited amount of data, the effectiveness of the adapted DNN model can often be compromised. We formulate maximum a posteriori (MAP) adaptation of parameters of a specially designed CD-DNN-HMM with an augmented linear hidden networks connected to the output tied states, or senones, and compare it to feature space MAP linear regression previously proposed. Experimental evidences on the 20,000-word open vocabulary Wall Street Journal task demonstrate the feasibility of the proposed framework. In supervised adaptation, the proposed MAP adaptation approach provides more than 10% relative error reduction and consistently outperforms the conventional transformation based methods. Furthermore, we present an initial attempt to generate hierarchical priors to improve adaptation efficiency and effectiveness with limited adaptation data by exploiting similarities among senones.

Speech recognition scenarios (aka tasks) differ from each other in acoustic transducers, acoustic environments, and speaking style etc. Building one acoustic model per task is one common practice in industry. However, this limits training data sharing across scenarios thus may not give highest possible accuracy. Based on the deep neural network (DNN) technique, we propose to build a universal acoustic model for all scenarios by utilizing all the data together. Two advantages are obtained: 1) leveraging more data sources to improve the recognition accuracy, 2) reducing substantially service deployment and maintenance costs. We achieve this by extending the singular value decomposition (SVD) structure of DNNs. The data from all scenarios are used to first train a single SVD-DNN model. Then a series of scenario-dependent linear square matrices are added on top of each SVD layer and updated with only scenario-related data. At the recognition time, a flag indicates different scenarios and guides the recognizer to use the scenario-dependent matrices together with the scenario-independent matrices in the universal DNN for acoustic score evaluation. In our experiments on Microsoft Winphone/Skype/Xbox data sets, the universal DNN model is better than traditional trained isolated models, with up to 15.5% relative word error rate reduction.

In this paper, we propose a novel method to adapt context-dependent deep neural network hidden Markov model (CD-DNN-HMM) with only limited number of parameters by taking into account the underlying factors that contribute to the distorted speech signal. We derive this factorized adaptation method from the perspectives of joint factor analysis and vector Taylor series expansion, respectively. Evaluated on Aurora 4, the proposed method can get 19.0% and 10.6% relative word error rate reduction on test set B and D with only 20 adaptation utterances, and can have decent improvement with as few as two adaptation utterances. We also show that the proposed method is better than feature discriminative linear regression (fDLR), an existing DNN adaptation method. Its small number of parameters and short training time offer an attractive solution to low-footprint speech applications.

We explore the use of maxout neuron in various aspects of acoustic modelling for large vocabulary speech recognition systems; including low-resource scenario and multilingual knowledge transfers. Through the experiments on voice search and short message dictation datasets, we found that maxout networks are around three times faster to train and offer lower or comparable word error rates on several tasks, when compared to the networks with logistic nonlinearity. We also present a detailed study of the maxout unit internal behaviour suggesting the use of different nonlinearities in different layers.

In this paper, we study a novel way to compensate speech features to counter the effects of speaker variations and environment distortions in speech recognition. For each homogeneous cluster of speech data, e.g. a specific speaker and environment combination, a set of correction vectors are learnt. A correction vector measures the deviation of features in a small region of feature space due to the speaker and environment effects. From a heterogenous training set, dozens of sets of correction vectors are learnt, each from a homogenous subset of the data. During testing, those correction vector sets are linearly combined to compensate test feature vectors. The combination weights are estimated by maximizing the likelihood (ML) of the compensated features with respect to a reference model, which is a simplified version of the acoustic model used for speech recognition. In addition, variance compensation is applied to condition the variances of the compensated features during weight estimation. Experimental results on Aurora-4 multi-condition training task show that the proposed correction vector combination method reduces the word error rate (WER) to 14.97% from mean and variance normalization baseline (16.32%) for noisy test sets 2-7. In addition, the proposed ML weight estimation consistently outperforms the posterior weights used in previous studies, such as multi-environment SPLICE.

The large number of parameters in deep neural networks (DNN) for automatic speech recognition (ASR) makes speaker adaptation very challenging. It also limits the use of speaker personalization due to the huge storage cost in large-scale deployments. In this paper we address DNN adaptation and personalization issues by presenting two methods based on the singular value decomposition (SVD). The first method uses an SVD to replace the weight matrix of a speaker independent DNN by the product of two low rank matrices. Adaptation is then performed by updating a square matrix inserted between the two low-rank matrices. In the second method, we adapt the full weight matrix but only store the delta matrix - the difference between the original and adapted weight matrices. We decrease the footprint of the adapted model by storing a reduced rank version of the delta matrix via an SVD. The proposed methods were evaluated on short message dictation task. Experimental results show that we can obtain similar accuracy improvements as the previously proposed Kullback-Leibler divergence (KLD) regularized method with far fewer parameters, which only requires 0.89% of the original model storage.

We propose a feature space maximum a posteriori (MAP) linear regression framework to adapt parameters for context dependent deep neural network hidden Markov models (CD-DNN-HMMs). Due to the huge amount of parameters used in DNN acoustic models in large vocabulary continuous speech recognition, the problem of over-fitting can be severe in DNN adaptation, thus often impair the robustness of the adapted DNN model. Linear input network (LIN) as a straight-forward feature space adaptation method for DNN, similar to feature space maximum likelihood linear regression (fMLLR), can potentially suffer from the same robustness situation. The proposed adaptation framework is built based on MAP estimation of the LIN parameters by incorporating prior knowledge into the adaptation process. Experimental results on the Switchboard task show that against the speaker independent CD-DNN-HMM systems, LIN provides 4.28% relative word error rate reduction (WERR) and the proposed fMAPLIN method is able to provide further 1.15% (totally 5.43%) WERR on top of LIN.

We propose two approaches for improving the objective function for the deep neural network (DNN) frame-level training in large vocabulary continuous speech recognition (LVCSR). The DNNs used in LVCSR are often constructed with an output layer with softmax activation and the cross-entropy objective function is always employed in the frame-leveling training of DNNs. The pairing of softmax activation and crossentropy objective function contributes much in the success of DNN. The first approach developed in this paper improves the cross-entropy objective function by boosting the importance of the frames for which the DNN model has low target predictions (low target posterior probabilities) and the second one considers jointly minimizing the cross-entropy and maximizing the log posterior ratio between the target senone (tied-triphone states) and the most competing one. Experiments on Switchboard task demonstrate that the two proposed methods can provide 3.1% and 1.5% relative word error rate (WER) reduction , respectively, against the already very strong conventional crossentropy trained DNN system.

Deep neural network (DNN) obtains significant accuracy improvements on many speech recognition tasks and its power comes from the deep and wide network structure with a very large number of parameters. It becomes challenging when we deploy DNN on devices which have limited computational and storage resources. The common practice is to train a DNN with a small number of hidden nodes and a small senone set using the standard training process, leading to significant accuracy loss. In this study, we propose to better address these issues by utilizing the DNN output distribution. To learn a DNN with small number of hidden nodes, we minimize the Kullback–Leibler divergence between the output distributions of the small-size DNN and a standard large-size DNN by utilizing a large number of un-transcribed data. For better senone set generation, we cluster the senones in the large set into a small one by directly relating the clustering process to DNN parameters, as opposed to decoupling the senone generation and DNN training process in the standard training. Evaluated on a short message dictation task, the proposed two methods get 5.08% and 1.33% relative word error rate reduction from the standard training method, respectively.

In a previous study, we proposed variable-component deep neural network (VCDNN) to improve the robustness of context-dependent deep neural network hidden Markov model (CD-DNN-HMM). We model the components of DNN a set of polynomial functions of environmental variables, more specifically signal-to-noise ratio (SNR). We refined VCDNN on two types of DNN components: (1) weighting matrix and bias (2) the output of each layer. These two methods are called variable-parameter DNN (VPDNN) and variable-output DNN (VODNN). Although both methods got good gain over the standard DNN, they doubled the number of parameters even with only the first-order environment variable. In this study, we propose two new types of VCDNN, namely variable activation DNN (VADNN) and variable input DNN (VIDNN). The environment variable is applied to the hidden layer activation function in VADNN, and is applied directly to the input in VIDNN. Both DNNs only increase a negligible number of parameters compared to the standard DNN. Experimental results on Aurora4 task show that both methods are effective, and VIDNN can beat all other variations of VCDNN with relative 7.69% word error reduction from the standard DNN with the least increase in number of parameters.

New waves of consumer-centric applications, such as voice search and voice interaction with mobile devices and home entertainment systems, increasingly require automatic speech recognition (ASR) to be robust to the full range of real-world noise and other acoustic distorting conditions. Despite its practical importance, however, the inherent links between and distinctions among the myriad of methods for noise-robust ASR have yet to be carefully studied in order to advance the field further. To this end, it is critical to establish a solid, consistent, and common mathematical foundation for noise-robust ASR, which is lacking at present. This article is intended to fill this gap and to provide a thorough overview of modern noise-robust techniques for ASR developed over the past 30 years. We emphasize methods that are proven to be successful and that are likely to sustain or expand their future applicability. We distill key insights from our comprehensive overview in this field and take a fresh look at a few old problems, which nevertheless are still highly relevant today. Specifically, we have analyzed and categorized a wide range of noise-robust techniques using five different criteria: 1) feature-domain vs. model-domain processing, 2) the use of prior knowledge about the acoustic environment distortion, 3) the use of explicit environment-distortion models, 4) deterministic vs. uncertainty processing, and 5) the use of acoustic models trained jointly with the same feature enhancement or model adaptation process used in the testing stage. With this taxonomy-oriented review, we equip the reader with the insight to choose among techniques and with the awareness of the performance-complexity tradeoffs. The pros and cons of using different noise-robust ASR techniques in practical application scenarios are provided as a guide to interested practitioners. The current challenges and future research directions in this field is also carefully analyzed.

Deep learning is becoming a mainstream technology for speech recognition at industrial scale. In this paper, we provide an overview of the work by Microsoft speech researchers since 2009 in this area, focusing on more recent advances which shed light to the basic capabilities and limitations of the current deep learning technology. We organize this overview along the feature-domain and model-domain dimensions according to the conventional approach to analyzing speech systems. Selected experimental results, including speech recognition and related applications such as spoken dialogue and language modeling, are presented to demonstrate and analyze the strengths and weaknesses of the techniques described in the paper. Potential improvement of these techniques and future research directions are discussed.

In the deep neural network (DNN), the hidden layers can be considered as increasingly complex feature transformations and the final softmax layer as a log-linear classifier making use of the most abstract features computed in the hidden layers. While the loglinear classifier should be different for different languages, the feature transformations can be shared across languages. In this paper we propose a shared-hidden-layer multilingual DNN (SHL-MDNN), in which the hidden layers are made common across many languages while the softmax layers are made language dependent. We demonstrate that the SHL-MDNN can reduce errors by 3-5%, relatively, for all the languages decodable with the SHL-MDNN, over the monolingual DNNs trained using only the language specific data. Further, we show that the learned hidden layers sharing across languages can be transferred to improve recognition accuracy of new languages, with relative error reductions ranging from 6% to 28% against DNNs trained without exploiting the transferred hidden layers. It is particularly interesting that the error reduction can be achieved for the target language that is in different families of the languages used to learn the hidden layers.

Context-dependent deep neural network HMMs have been shown to achieve recognition accuracy superior to Gaussian mixture models in a number of recent works.Typically, neural networks are optimized with stochastic gradient descent.On large datasets, stochastic gradient descent improves quickly during the beginning of the optimization.But since it does not make use of second order information, its asymptotic convergence behavior is slow.In regions with pathological curvature, stochastic gradient descent may almost stagnate and thereby falsely indicate convergence.Another drawback of stochastic gradient descent is that it can only be parallelized within minibatches.The Hessian-free algorithm is a second order batch optimization algorithm that does not suffer from these problems.In a recent work, Hessian-free optimization has been applied to a training of deep neural networks according to a sequence criterion.In that work, improvements in accuracy and training time have been reported.In this paper, we analyze the properties of the Hessian-free optimization algorithm and investigate whether it is suited for cross-entropy training of deep neural networks as well.

Recently proposed deep neural network (DNN) obtains significant accuracy improvements in many large vocabulary continuous speech recognition (LVCSR) tasks. However, DNN requires much more parameters than traditional systems, which brings huge cost during online evaluation, and also limits the application of DNN in a lot of scenarios. In this paper we present our new effort on DNN aiming at reducing the model size while keeping the accuracy improvements. We apply singular value decomposition (SVD) on the weight matrices in DNN, and then restructure the model based on the inherent sparseness of the original matrices. After restructuring we can reduce the DNN model size significantly with negligible accuracy loss. We also fine-tune the restructured model using the regular back-propagation method to get the accuracy back when reducing the DNN model size heavily. The proposed method has been evaluated on two LVCSR tasks, with context-dependent DNN hidden Markov model (CD-DNN-HMM). Experimental results show that the proposed approach dramatically reduces the DNN model size by more than 80% without losing any accuracy. Index Terms: deep neural network, singular value decomposition, model restructuring

Recently, speech scientists have been motivated by the great, success of building margin‐based classifiers, and have thus proposed novel methods to estimate continuous‐density hidden Markov model (HMM) for automatic speech recognition (ASR) according to the notion that the decision boundaries determined by the estimated HMMs attain the maximum classification margin as in learning support vector machines. Although a good performance has been observed, the margin used in the ASR community is often specified as a parameter that has no explicit relationship with the HMM parameters. The issues of how the margin is related to the HMM parameters and how it directly characterises the generalisation capability of HMM‐based classifiers have not been addressed so far in the community. In this study, the authors attempt to formulate the margin used in the soft margin estimation framework as a function of the HMM parameters. The key idea is to relate the standard distance‐based margin with the concept of divergence among competing HMM state Gaussian mixture model densities. Experimental results show that the proposed model‐based margin function is a good indication about the quality of HMMs on a given ASR task without the conventional needs of running experiments extensively using a separate set of test samples.

Model adaptation techniques are an efficient way to reduce the mismatch that typically occurs between the training and test condition of any automatic speech recognition (ASR) system. This work addresses the problem of increased degradation in performance when moving from speaker-dependent (SD) to speaker-independent (SI) conditions for connectionist (or hybrid) hidden Markov model/artificial neural network (HMM/ANN) systems in the context of large vocabulary continuous speech recognition (LVCSR). Adapting hybrid HMM/ANN systems on a small amount of adaptation data has been proven to be a difficult task, and has been a limiting factor in the widespread deployment of hybrid techniques in operational ASR systems. Addressing the crucial issue of speaker adaptation (SA) for hybrid HMM/ANN system can thereby have a great impact on the connectionist paradigm, which will play a major role in the design of next-generation LVCSR considering the great success reported by deep neural networks - ANNs with many hidden layers that adopts the pre-training technique - on many speech tasks. Current adaptation techniques for ANNs based on injecting an adaptable linear transformation network connected to either the input, or the output layer are not effective especially with a small amount of adaptation data, e.g., a single adaptation utterance. In this paper, a novel solution is proposed to overcome those limits and make it robust to scarce adaptation resources. The key idea is to adapt the hidden activation functions rather than the network weights. The adoption of Hermitian activation functions makes this possible. Experimental results on an LVCSR task demonstrate the effectiveness of the proposed approach.

By explicitly modelling the distortion of speech signals, model adaptation based on vector Taylor series (VTS) approaches have been shown to significantly improve the robustness of speech recognizers to environmental noise. However, the computational cost of VTS model adaptation (MVTS) methods hinders them from being widely used because they need to adapt all the HMM parameters for every utterance at runtime. In contrast, VTS feature enhancement (FVTS) methods have more computation advantages because they do not need multiple decoding passes and do not adapt all the HMM model parameters. In this paper, we propose two improvements to VTS feature enhancement: updating all of the environment distortion parameters and noise adaptive training of the front-end GMM. In addition, we investigate some other performance-related issues such as the selection of FVTS algorithms and the spectrum domain that MFCC is extracted from. As an important result of our investigation, we established the FVTS method can achieve comparable accuracy as the MVTS method with a smaller runtime cost. This makes FVTS method an ideal candidate for real world tasks.

In this paper, we propose a novel acoustic model adaptation method for noise robust speech recognition. Model combination is a common way to adapt acoustic models to a target test environment. For example, the mean supervectors of the adapted model are obtained as a linear combination of mean supervectors of many pre-trained environment-dependent acoustic models. Usually, the combination weights are estimated using a maximum likelihood (ML) criterion and the weights are nonzero for all the mean supervectors. We propose to estimate the weights by using Lasso (least absolute shrinkage and selection operator) which imposes an L1regularization term in the weight estimation problem to shrink some weights to exactly zero. Our study shows that Lasso usually shrinks to zero the weights of those mean supervectors not relevant to the test environment. By removing some nonrelevant supervectors, the obtained mean supervectors are found to be more robust against noise distortions. Experimental results on Aurora-2 task show that the Lasso-based mean combination consistently outperforms ML-based combination.

By exploiting a model of environmental distortion, model adaptation based on vector Taylor series (VTS) approaches have been shown to significantly improve the robustness of speech recognizers to environmental noise. However, the computational cost of VTS model adaptation (MVTS) methods hinders them from being more widely used. In this paper, we propose to reduce the computational cost of MVTS by replacing the Jacobian matrix used in the vector Taylor series approximation with a diagonal Jacobian matrix (DJVTS). We verify this approximation by showing that the Jacobian matrices are dominated by their diagonal elements and therefore the model distortion introduced by this approximation is very small. DJVTS gives similar accuracy as the standard MVTS method with significant reduction in computational cost. The proposed method also achieves higher accuracy than VTS-based feature enhancement.

This work is concerned with speaker adaptation techniques for artificial neural network (ANN) implemented as feed-forward multi-layer perceptrons (MLPs) in the context of large vocabulary continuous speech recognition (LVCSR). Most successful speaker adaptation techniques for MLPs consist of augmenting the neural architecture with a linear transformation network connected to either the input or the output layer. The weights of this additional linear layer are learned during the adaptation phase while all of the other weights are kept frozen in order to avoid over-fitting. In doing so, the structure of the speaker-dependent (SD) and speaker-independent (Si) architecture differs and the number of adaptation parameters depends upon the dimension of either the input or output layers. We propose an alternative neural architecture for speaker-adaptation to overcome the limits of current approaches. This neural architecture adopts hidden activation functions that can be learned directly from the adaptation data. This adaptive capability of the hidden activation function is achieved through the use of orthonormal Hermite polynomials. Experimental evidence gathered on the Wall Street Journal Nov92 task demonstrates the viability of the proposed technique.

Context-dependent deep neural network hidden Markov model (CD-DNN-HMM) is a recently proposed acoustic model that significantly outperformed Gaussian mixture model (GMM)-HMM systems in many large vocabulary speech recognition (LVSR) tasks. In this paper we present our strategy of using mixed-bandwidth training data to improve wideband speech recognition accuracy in the CD-DNN-HMM framework. We show that DNNs provide the flexibility of using arbitrary features. By using the Mel-scale log-filter bank features we not only achieve higher recognition accuracy than using MFCCs, but also can formulate the mixed-bandwidth training problem as a missing feature problem, in which several feature dimensions have no value when narrowband speech is presented. This treatment makes training CD-DNN-HMMs with mixed-bandwidth data an easy task since no bandwidth extension is needed. Our experiments on voice search data indicate that the proposed solution not only provides higher recognition accuracy for the wideband speech but also allows the same CD-DNN-HMM to recognize mixed-bandwidth speech. By exploiting mixed-bandwidth training data CD-DNN-HMM outperforms fMPE+BMMI trained GMM-HMM, which cannot benefit from using narrowband data, by 18.4%.

In this paper, we propose a novel feature space adaptation technique to improve the robustness of speech recognition in noisy environments. Histogram equalization (HEQ) is an effective technique for improving robustness by reducing the difference between clean and noisy features. A weakness of HEQ is that it does not take into account acoustic model, resulting in possible mismatch between HEQ processed features and the acoustic model. In this paper, we propose to adapt HEQ to maximize the likelihood of HEQ-processed features on the acoustic model, with a constraint on the parameters of HEQ. In addition, we use a Gaussian mixture model (GMM) to represent the clean feature space rather than using the acoustic model itself, and this results in both simpler implementation and better results. Experimental results show that HEQ with adaptation reduces word error rate by 7.5% and 5.7% respectively on Aurora-2 and Aurora-4 tasks over the HEQ baseline without adaptation.

Classical mean and variance normalization (MVN) uses a di- agonal transform and a bias vector to normalize the mean and variance of noisy features to reference values. As MVN uses di- agonal transform, it ignores correlation between feature dimen- sions. Although full transform is able to make use of feature correlation, its large amount of parameters may not be estimated reliably from a short observation, e.g. 1 utterance. We propose a novel structured full transform that has the same amount of free parameters as diagonal transform while being able to capture correlation between feature dimensions. The proposed struc- tured transform can be estimated reliably from one utterance by maximizing the likelihood of the normalized features on a refer- ence Gaussian mixture model. Experimental results on Aurora- 4 task show that the structured transform produces consistently better speech recognition results than diagonal transform and also outperforms advanced frontend (AFE) feature extractor.

Most speech recognition applications in use today rely heavily on confidence measure for making optimal decisions. In this paper, we aim to answer the question: what can be done to improve the quality of confidence measure if we cannot modify the speech recognition engine? The answer provided in this paper is a post-processing step called confidence calibration, which can be viewed as a special adaptation technique applied to confidence measure. Three confidence calibration methods have been developed in this work: the maximum entropy model with distribution constraints, the artificial neural network, and the deep belief network. We compare these approaches and demonstrate the importance of key features exploited: the generic confidence-score, the application-dependent word distribution, and the rule coverage ratio. We demonstrate the effectiveness of confidence calibration on a variety of tasks with significant normalized cross entropy increase and equal error rate reduction.

It is widely known that the quality of confidence measure is critical for speech applications. In this paper, we present our recent work on improving word confidence scores by calibrating them using a small set of calibration data when only the recognized word sequence and associated raw confidence scores are made available. The core of our technique is the maximum entropy model with distribution constraints which naturally and effectively make use of the word distribution, the raw confidence-score distribution, and the context information. We demonstrate the effectiveness of our approach by showing that it can achieve relative 38% mean square error (MSE), 39% negative normalized likelihood (NNLL), and 23% equal error rate (EER) reduction on a voice mail transcription data set and relative 35% MSE, 45% NNLL, and 35% EER reduction on a command and control data set.

Inspired by the success of least absolute shrinkage and selection operator (LASSO) in statistical learning, we propose an regularized maximum likelihood linear regression (MLLR) to estimate models with only a limited set of adaptation data to improve accuracy for automatic speech recognition, by regularizing the standard MLLR objective function with an constraint. The so-called LASSO MLLR is a natural solution to the data insufficiency problem because the constraint regularizes some parameters to exactly 0 and reduces the number of free parameters to estimate. Tested on the 5k-WSJ0 task, the proposed LASSO MLLR gives significant word error rate reduction from the errors obtained with the standard MLLR in an utterance-by-utterance unsupervised adaptation scenario. 1.

In this paper, we propose to improve our previously developed method for joint compensation of additive and convolutive distortions (JAC) applied to model adaptation. The improvement entails replacing the vector Taylor series (VTS) approximation with unscented transform (UT) in formulating both the static and dynamic model parameter adaptation. Our new JAC-UT method differentiates itself from other UT-based approaches in that it combines the online noise and channel distortion estimation and model parameter adaptation in a unified UT framework. Experimental results on the standard Aurora 2 task show that the new algorithm enjoys 20.0 % and 16.9 % relative word error rate reductions over the previous JAC-VTS algorithm when using the simple and complex backend models, respectively. Index Terms: unscented transform, vector Taylor series, additive and convolutive distortions, robust ASR, adaptation

In this paper, we explore the generalization capability of acoustic model for improving speech recognition robustness against noise distortions. While generalization in statistical learning theory originally refers to the model's ability to generalize well on unseen testing data drawn from the same distribution as that of the training data, we show that good generalization capability is also desirable for mismatched cases. One way to obtain such general models is to use margin-based model training method, e.g., soft-margin estimation (SME), to enable some tolerance to acoustic mismatches without a detailed knowledge about the distortion mechanisms through enhancing margins between competing models. Experimental results on the Aurora-2 and Aurora-3 connected digit string recognition tasks demonstrate that, by improving the model's generalization capability through SME training, speech recognition performance can be significantly improved in both matched and low to medium mismatched testing cases with no language model constraints. Recognition results show that SME indeed performs better with than without mean and variance normalization, and therefore provides a complimentary benefit to conventional feature normalization techniques such that they can be combined to further improve the system performance. Although this study is focused on noisy speech recognition, we believe the proposed margin-based learning framework can be extended to dealing with different types of distortions and robustness issues in other machine learning applications.

From statistical learning theory, the generalization capability of a model is the ability to generalize well on unseen test data which follow the same distribution as the training data. This paper investigates how generalization capability can also improve robustness when testing and training data are from different distributions in the context of speech recognition. Two discriminative training (DT) methods are used to train the hidden Markov model (HMM) for better generalization capability, namely the minimum classification error (MCE) and the soft-margin estimation (SME) methods. Results on Aurora-2 task show that both SME and MCE are effective in improving one of the measures of acoustic model's generalization capability, i.e. the margin of the model, with SME be moderately more effective. In addition, the better generalization capability translates into better robustness of speech recognition performance, even when there is significant mismatch between the training and testing data. We also applied the mean and variance normalization (MVN) to preprocess the data to reduce the training-testing mismatch. After MVN, MCE and SME perform even better as the generalization capability now is more closely related to robustness. The best performance on Aurora-2 is obtained from SME and about 28% relative error rate reduction is achieved over the MVN baseline system. Finally, we also use SME to demonstrate the potential of better generalization capability in improving robustness in more realistic noisy task using the Aurora-3 task, and significant improvements are obtained.

Recently, we proposed an ensemble speaker and speaking environment modeling (ESSEM) framework to characterize speaker variability and speaking environments. In contrast to multi-style training, ESSEM uses single-style training to prepare multiple sets of environment-specific acoustic models. The ensemble of these acoustic models forms a prior structure of the environment for flexible prediction of unknown environment during testing. In this study, we present methods to further improve the precision for model characterization. We first study a weighted N-best information technique to well utilize the N-best transcription hypothesis in an unsupervised adaptation manner. Next, we introduce cohort selection and environment space adaptation techniques to online improve the resolution and coverage of the prior structure. With an integration of the proposed methods, we further improve the ESSEM performance over our previous study. On the Aurora-2 task, ESSEM achieves an average word error rate (WER) of 4.64%, corresponding to a 15.64% relative WER reduction over our best baseline result (5.50% to 4.64% WER) obtained with multi-condition training.

We formulate a framework for soft margin estimation-based linear regression (SMELR) and apply it to supervised speaker adaptation. Enhanced separation capability and increased discriminative ability are two key properties in margin-based discriminative training. For the adaptation process to be able to flexibly utilize any amount of data, we also propose a novel interpolation scheme to linearly combine the speaker independent (SI) and speaker adaptive SMELR (SMELR/SA) models. The two proposed SMELR algorithms were evaluated on a Japanese large vocabulary continuous speech recognition task. Both the SMELR and interpolated SI+SMELR/SA techniques showed improved speech adaptation performance in comparison with the well-known maximum likelihood linear regression (MLLR) method. We also found that the interpolation framework works even more effectively than SMELR when the amount of adaptation data is relatively small.

In this paper, we present a new approach to HMM adaptation that jointly compensates for additive and convolutive acoustic distortion in environment-robust speech recognition. The hallmark of our new approach is the use of a nonlinear, phase-sensitive model of acoustic distortion that captures phase asynchrony between clean speech and the mixing noise. In the first step of the developed algorithm, both the static and dynamic portions of the noise and channel parameters are estimated in the cepstral domain, using the speech recognizer’s “feedback” information and the vector-Taylor-series linearization technique on the nonlinear phase-sensitive model. In the second step, the estimated noise and channel parameters are used to effectively adapt the static and dynamic portions of the HMM means and variances also using the linearized phase-sensitive acoustic distortion model. In the experimental evaluation using the standard Aurora 2 task, the proposed new algorithm achieves 93.3% accuracy using the clean-trained complex HMM backend as the baseline system for unsupervised HMM adaptation. This reaches the highest performance number in the literature on this task with clean-trained HMM model. The experimental results show that the phase term, which was missing in all previous HMM-adaptation work, contributes significantly to the achieved high recognition accuracy.

Recently, we successfully developed and reported a new unsupervised online adaptation technique, which jointly compensates for additive and convolutive distortions with vector Taylor series (JAC/VTS), to adjust (uncompressed) HMMs under acoustically distorted environments [1]. In this paper, we extend that technique to adapt compressed HMMs using JAC/VTS where limited computation and/or memory resources are available for speech recognition (e.g., on mobile devices). Subspace coding (SSC) is developed and used to quantize each dimension of the multivariate Gaussians in the compressed HMMs. Three algorithmic design options are proposed and evaluated that combine SSC with JAC/VTS, where three different types of tradeoffs are made between recognition accuracy and the required computation/memory/storage resources. The strengths and weaknesses of these three options are discussed and shown on the Aurora2 task of noise-robust speech recognition. The first option greatly reduces the storage space and gives 93.2% accuracy, which is the same as the baseline accuracy but with little reduction in the run-time computation/memory cost. The second option reduces about 79.9% of the computation cost and about 33.5% of the memory requirement at a very small price of 0.5% decrease of accuracy (to 92.7%). The third option cuts about 89.2% of the computation cost and about 65.5% of the memory requirement while reducing recognition accuracy by 2.7% (to 90.5%).

Recently, there have been intensive studies of margin-based learning for automatic speech recognition (ASR). It is our believe that by securing a margin from the decision boundaries to the training samples, a correct decision can still be made if the mismatches between testing and training samples are well within the tolerance region specified by the margin. This nice property should be effective for robust ASR, where the testing condition is different from those in training. In this paper, we report on experiment results with soft margin estimation (SME) on the Aurora2 task and show that SME is very effective under clean training with more than 50% relative word error reductions in the clean, 20db, and 15db testing conditions, and still gives a slight improvement over conventional multi-condition training approaches. This demonstrates that the margin in SME can equip recognizers with a nice generalization property under adverse conditions.

We continue our previous work on soft margin estimation (SME) to large vocabulary continuous speech recognition (LVCSR) in two new aspects. The first is to formulate SME with different unit separation. SME methods focusing on string-, word-, and phone-level separation are defined. The second is to compare SME with all the popular conventional discriminative training (DT) methods, including maximum mutual information estimation (MMIE), minimum classification error (MCE), and minimum word/phone error (MWE/MPE). Tested on the 5k-word Wall Street Journal task, all the SME methods achieves a relative word error rate (WER) reduction from 17% to 25% over our baseline. Among them, phone-level SME obtains the best performance. Its performance is slightly better than that of MPE, and much better than those of other conventional DT methods. With the comprehensive comparison with conventional DT methods, SME demonstrates its success on LVCSR tasks.

In this paper, we present our recent development of a model-domain environment-robust adaptation algorithm, which demonstrates high performance in the standard Aurora 2 speech recognition task. The algorithm consists of two main steps. First, the noise and channel parameters are estimated using a nonlinear environment distortion model in the cepstral domain, the speech recognizer’s “feedback” information, and the Vector-Taylor-Series (VTS) linearization technique collectively. Second, the estimated noise and channel parameters are used to adapt the static and dynamic portions of the HMM means and variances. This two-step algorithm enables Joint compensation of both Additive and Convolutive distortions (JAC). In the experimental evaluation using the standard Aurora 2 task, the proposed JAC/VTS algorithm achieves 91.11% accuracy using the clean-trained simple HMM backend as the baseline system for the model adaptation. This represents high recognition performance on this task without discriminative training of the HMM system. Detailed analysis on the experimental results shows that adaptation of the dynamic portion of the HMM mean and variance parameters is critical to the success of our algorithm.

We extend our previous work on soft margin estimation (SME) to large vocabulary continuous speech recognition in two aspects. The first is to use the extended Baum-Welch method to replace the conventional generalized probabilistic descent algorithm for optimization. The second is to compare SME with minimum classification error (MCE) training with the same implementation details in order to show that it is indeed the margin component in the objective function with margin-based utterance and frame selection that contributes to the success of SME. Tested on the 5 k-word Wall Street Journal task, all the SME methods work better than MCE. The best SME approach achieves a relative word error rate reduction of about 19% over our best baseline performance. This enhancement can only be demonstrated because of our use of margin-based objective function and the extended Baum-Welch parameter optimization method.

In a recent study, we proposed soft margin estimation (SME) to learn parameters of continuous density hidden Markov models (HMMs). Our earlier experiments with connect digit recognition have shown that SME offers great advantages over other state-of-the-art discriminative training methods. In this paper, we illustrate SME from a perspective of statistical learning theory and show that by including a margin in formulating the SME objective function it is capable of directly minimizing the approximate test risk, while most other training methods intent to minimize only the empirical risks. We test SME on the 5k-word Wall Street Journal task, and find the proposed approach achieves a relative word error rate reduction of about 10% over our best baseline results in different experimental configurations. We believe this is the first attempt to show the effectiveness of margin-based acoustic modeling for large vocabulary continuous speech recognition. We also expect further performance improvements in the future because the approximate test risk minimization principle offers a flexible and yet rigorous framework to facilitate easy incorporation of new margin-based optimization criteria into HMM training.

We present methods of detector design in the Automatic Speech Attribute Transcription project. This paper details the results of a student-led, cross-site collaboration between Georgia Institute of Technology, The Ohio State University and Rutgers University. The work reported in this paper describes and evaluates the detection-based ASR paradigm and discusses phonetic attribute classes, methods of detecting framewise phonetic attributes and methods of combining attribute detectors for ASR. We use Multi-Layer Perceptrons, Hidden Markov Models and Support Vector Machines to compute confidence scores for several prescribed sets of phonetic attribute classes. We use Conditional Random Fields (CRFs) and knowledge-based rescoring of phone lattices to combine framewise detection scores for continuous phone recognition on the TIMIT database. With CRFs, we achieve a phone accuracy of 70.63%, outperforming the baseline and enhanced HMM systems, by incorporating all of the attribute detectors discussed in the paper

We propose a new discriminative learning framework, called soft margin feature extraction (SMFE), for jointly optimizing the parameters of transformation matrix for feature extraction and of hidden Markov models (HMMs) for acoustic modeling. SMFE extends our previous work of soft margin estimation (SME) to feature extraction. Tested on the TIDIGITS connected digit recognition task, the proposed approach achieves a string accuracy of 99.61%, much better than our previously reported SME results. To our knowledge, this is the first study on applying the margin-based method in joint optimization of feature extraction and acoustic modeling. The excellent performance of SMFE demonstrates the success of soft margin based method, which targets to obtain both high accuracy and good model generalization. Index Terms: discriminative feature extraction, hidden Markov model, margin, automatic speech recognition

Inspired by the great success of margin-based classifiers, there is a trend to incorporate the margin concept into hidden Markov modeling for speech recognition. Several attempts based on margin maximization were proposed recently. In this paper, a new discriminative learning framework, called soft margin estimation (SME), is proposed for estimating the parameters of continuous-density hidden Markov models. The proposed method makes direct use of the successful ideas of soft margin in support vector machines to improve generalization capability and decision feedback learning in minimum classification error training to enhance model separation in classifier design. SME is illustrated from a perspective of statistical learning theory. By including a margin in formulating the SME objective function, SME is capable of directly minimizing an approximate test risk bound. Frame selection, utterance selection, and discriminative separation are unified into a single objective function that can be optimized using the generalized probabilistic descent algorithm. Tested on the TIDIGITS connected digit recognition task, the proposed SME approach achieves a string accuracy of 99.43%. On the 5 k-word Wall Street Journal task, SME obtains relative word error rate reductions of about 10% over our best baseline results in different experimental configurations. We believe this is the first attempt to show the effectiveness of margin-based acoustic modeling for large vocabulary continuous speech recognition in a hidden Markov model framework. Further improvements are expected because the approximate test risk bound minimization principle offers a flexible and rigorous framework to facilitate incorporation of new margin-based optimization criteria into hidden Markov model training.

We propose a new discriminative learning framework, called soft margin estimation (SME), for estimating parameters of continuous density hidden Markov models. The proposed method makes direct usage of the successful ideas of soft margin in support vector machines to improve generalization capability, and of decision feedback learning in minimum classification error training to enhance model separation in classifier design. We attempt to incorporate frame selection, utterance selection and discriminative separation in a single unified objective function that can be optimized with the wellknown generalized probabilistic descent algorithm. We demonstrate the advantage of SME in theory and practice over other state-of-the-art techniques. Tested on a connected digit recognition task, the proposed SME approach achieves a string accuracy of 99.33%. To our knowledge, this is the best result ever reported on the TIDIGITS database. 1.

We study lattice rescoring with knowledge scores for automatic speech recognition. Frame-based log likelihood ratio is adopted as a score measure of the goodness-of-fit between a speech segment and the knowledge sources. We evaluate our approach in two different applications: phone recognition, and connected digit continuous recognition. By incorporating knowledge scores obtained from 15 attribute detectors for place and manner of articulation, we reduced phone error rate from 40.52% to 35.16% using monophone models. The error rate can be further reduced to 33.42% for triphone models. The same lattice rescoring algorithm is extended to connected digit recognition using the TIDIGITS database, and without using any digit-specific training data. We observed the digit error rate can be effectively reduced to 4.03% from 4.54% which was obtained with the conventional Viterbi decoding algorithm with no knowledge scores.

We propose a rescoring framework for speech recognition that incorporates acoustic phonetic knowledge sources. The scores corresponding to all knowledge sources are generated from a collection of neural network based classifiers. Rescoring is then performed by combining different knowledge scores and they are used to reorder candidate strings provided by state-of-the-art HMM-based speech recognizers. We report on continuous phone recognition experiments using the TIMIT database. Our results indicate that classifying manners and places of articulation provides additional information in rescoring, and improved accuracies over our best baseline speech recognizers are achieved using both context-independent and context-dependent phone models. The same technique can be extended to lattice rescoring and large vocabulary continuous speech recognition.

We study issues related to designing speech event detectors for automatic speech recognition. Event detection is a critical component of a recently proposed automatic speech attribute transcription (ASAT) paradigm for speech research. Similar to keyword spotting and non-keyword rejection, a good detector needs to effectively detect speech attributes of interest while rejecting extraneous events. We compare frame and segment based detectors, study their properties in detecting manners of articulation, and propose new performance measures. We test these detectors on the TIMIT database with several evaluation criteria. Our results indicate that segment based detectors outperform frame based detectors in several key aspects of speech detector design. We also show that the performance can be significantly enhanced by incorporating discriminative training into designing speech event detectors.

We study separation between models of speech attributes. A good measure of separation usually serves as a key indicator of the discrimination power of these speech models because it can often be used to indirectly determine the performance of speech recognition and verification systems. In this study, we use a probabilistic distance, called generalized log likelihood ratio (GLLR), to measure the separation between a model of a target speech attribute and models of its competing attributes. We illustrate five applications to compare separations among models obtained over multiple levels of discrimination capabilities, at various degrees of acoustic definitions and resolutions, under mismatched training and testing conditions, and with different training criteria and speech parameters. We demonstrate that the well-known GLLR distance and its corresponding histograms also provide a good utility to qualitatively and quantitatively characterize the properties of trained models without performing large scale speech recognition and verification experiments.

In this paper, the minimum classification error (MCE) method is extended to optimize both linear discriminant analysis (LDA) transformation and the classification parameters for dimensionality reduction. Firstly, under the HMM-based continuous speech recognition (CSR) framework, we use the MCE criterion to optimize the conventional dimensionality reduction method, which uses LDA to transform the standard MFCC. Then, a new dimensionality reduction method is proposed. In the new method, the combination of discrete cosine transform (DCT) and LDA, as used in the conventional method, is replaced by a single LDA transformation, which is optimized according to MCE criterion along with the classification parameters. Experimental results on TiDigits show that even when the feature dimension is reduced to 14, the performance of this new method is as good as that of the MCE-trained system using 39 dimension MFCC. It also outperforms our MCE-optimized conventional dimensionality reduction method.

In October 2002, the advanced front-end (AFE) for distributed speech recognition (DSR) was standardized by ETSI. In order to use the AFE feature on low computational resource devices, we propose a novel approach to improve the computational efficiency. In our new algorithm, the structure of the two-stage mel-warped Wiener filtering algorithm, which is the main part of AFE, is modified. A Wiener filter is constructed and applied directly in the mel-warped filter-bank domain. The measures we take make many time-consuming operations in the original algorithm completely unnecessary, including the re-calculations of power spectrum and the time-domain convolution operations. Consequently, a large amount of computations are saved. Experiments show that the new approach can substantially reduce the computation load while preserving the excellent performance of the ETSI AFE.

Because classical fast vector quantization(VQ) algorithms can’t be used in the LSF vector quantizers that use varying weighted Euclidean distance, a novel fast VQ search algorithm —CRVQCS (Constrained Range Vector Quantization based on Component Searching) is presented in this paper. The CRVQ-CS algorithm works well with the varying weighted Euclidean distance and yields the same result as full search VQ with reduced computational complexity does. Although the CRVQ-CS algorithm is proposed for VQ using varying weighted Euclidean distance measure, it is also suitable for VQ using simple Euclidean distance measure.