Yashesh Gaur

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Jian Xue, Peidong Wang, Jinyu Li et al. · microsoft-research

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.

Jian Wu, Yashesh Gaur, Zhuo Chen et al.

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.

Peidong Wang, Eric Sun, Jian Xue et al.

Automatic speech recognition (ASR) and speech translation (ST) can both use neural transducers as the model structure. It is thus possible to use a single transducer model to perform both tasks. In real-world applications, such joint ASR and ST models may need to be streaming and do not require source language identification (i.e. language-agnostic). In this paper, we propose LAMASSU, a streaming language-agnostic multilingual speech recognition and translation model using neural transducers. Based on the transducer model structure, we propose four methods, a unified joint and prediction network for multilingual output, a clustered multilingual encoder, target language identification for encoder, and connectionist temporal classification regularization. Experimental results show that LAMASSU not only drastically reduces the model size but also reaches the performances of monolingual ASR and bilingual ST models.

Jing Pan, Jian Wu, Yashesh Gaur et al.

We present a cost-effective method to integrate speech into a large language model (LLM), resulting in a Contextual Speech Model with Instruction-following/in-context-learning Capabilities (COSMIC) multi-modal LLM. Using GPT-3.5, we generate Speech Comprehension Test Question-Answer (SQA) pairs from speech transcriptions for supervised instruction tuning. With under 30 million trainable parameters and only 450 hours of English speech data, COSMIC demonstrates emerging capabilities in instruction-following and in-context learning. Equipped with such capabilities, COSMIC achieves a maximum 33.18 BLEU score in 0-shot EN-to-X speech to text translation (S2TT) and a significant boost in the 1-shot setting. Additionally, there is an average 25.8\% relative Word Error Rate (WER) reduction for 1-shot cross-domain adaptation. COSMIC exhibits a significant automatic speech recognition (ASR) accuracy gain in contextual biasing tasks due to its instruction-following capability.

Sara Papi, Peidong Wang, Junkun Chen et al.

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 1.3s, 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.

Yashesh Gaur, Nick Kibre, Jian Xue et al.

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 baselines, while being significantly smaller in size and retaining customization capabilities.

Sara Papi, Peidong Wang, Junkun Chen et al.

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.

Yi-Jen Shih, Desh Raj, Chunyang Wu et al.

Recent advances in speech large language models (speech LLMs) have enabled seamless spoken interactions, but these systems still struggle with complex reasoning tasks. Previously, chain-of-thought (CoT) prompting or fine-tuning has been to shown to significantly improve the reasoning abilities of text-based LLMs. In this work, we investigate the effect of CoT fine-tuning for multi-stream speech LLMs, demonstrating that reasoning in text space improves the accuracy of speech LLMs by 2.4x, on average, over a suite of spoken reasoning tasks. Beyond accuracy, the latency of the spoken response is a crucial factor for interacting with voice-based agents. Inspired by the human behavior of "thinking while listening," we propose methods to reduce the additional latency from reasoning by allowing the model to start reasoning before the user query has ended. To achieve this, we introduce an entropy-based metric, "question completeness," which acts as an indicator to guide the model on the optimal time to start reasoning. This method provides greater control over the accuracy-latency trade-off compared with heuristic-based approaches and, under equivalent latency conditions, yields a 4% accuracy gain on ARC-Easy. Finally, we use Direct Preference Optimization (DPO) on preference data created using rejection sampling to push the accuracy-latency pareto frontier further, resulting in a 70% reduction in latency without loss in accuracy.

Niko Moritz, Ruiming Xie, Yashesh Gaur et al.

We propose the joint speech translation and recognition (JSTAR) model that leverages the fast-slow cascaded encoder architecture for simultaneous end-to-end automatic speech recognition (ASR) and speech translation (ST). The model is transducer-based and uses a multi-objective training strategy that optimizes both ASR and ST objectives simultaneously. This allows JSTAR to produce high-quality streaming ASR and ST results. We apply JSTAR in a bilingual conversational speech setting with smart-glasses, where the model is also trained to distinguish speech from different directions corresponding to the wearer and a conversational partner. Different model pre-training strategies are studied to further improve results, including training of a transducer-based streaming machine translation (MT) model for the first time and applying it for parameter initialization of JSTAR. We demonstrate superior performances of JSTAR compared to a strong cascaded ST model in both BLEU scores and latency.

Yen-Ju Lu, Yashesh Gaur, Wei Zhou et al.

Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.

Frank Seide, Morrie Doulaty, Yangyang Shi et al.

We introduce Speech ReaLLM, a new ASR architecture that marries "decoder-only" ASR with the RNN-T to make multimodal LLM architectures capable of real-time streaming. This is the first "decoder-only" ASR architecture designed to handle continuous audio without explicit end-pointing. Speech ReaLLM is a special case of the more general ReaLLM ("real-time LLM") approach, also introduced here for the first time. The idea is inspired by RNN-T: Instead of generating a response only at the end of a user prompt, generate after every input token received in real time (it is often empty). On Librispeech "test", an 80M Speech ReaLLM achieves WERs of 3.0% and 7.4% in real time (without an external LM or auxiliary loss). This is only slightly above a 3x larger Attention-Encoder-Decoder baseline. We also show that this way, an LLM architecture can learn to represent and reproduce the flow of time; and that a pre-trained 7B LLM can be fine-tuned to do reasonably well on this task.

Tianrui Wang, Long Zhou, Ziqiang Zhang et al.

Recent research shows a big convergence in model architecture, training objectives, and inference methods across various tasks for different modalities. In this paper, we propose VioLA, 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 codec language model task via multi-task learning framework. To accomplish this, we first convert all 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 conversion problems, which can be naturally handled with one conditional language model. We further integrate task IDs (TID) and language IDs (LID) into the proposed model to enhance the modeling capability of handling different languages and tasks. Experimental results demonstrate that the proposed VioLA model 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.

Naoyuki Kanda, Jian Wu, Yu Wu et al.

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 encoder-decoder 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.

Naoyuki Kanda, Jian Wu, Yu Wu et al.

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.

Kenichi Kumatani, Dimitrios Dimitriadis, Yashesh Gaur et al.

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.

Naoyuki Kanda, Xiong Xiao, Yashesh Gaur et al.

This paper presents Transcribe-to-Diarize, a new approach for neural speaker diarization that uses an end-to-end (E2E) speaker-attributed automatic speech recognition (SA-ASR). The E2E SA-ASR is a joint model that was recently proposed for speaker counting, multi-talker speech recognition, and speaker identification from monaural audio that contains overlapping speech. Although the E2E SA-ASR model originally does not estimate any time-related information, we show that the start and end times of each word can be estimated with sufficient accuracy from the internal state of the E2E SA-ASR by adding a small number of learnable parameters. Similar to the target-speaker voice activity detection (TS-VAD)-based diarization method, the E2E SA-ASR model is applied to estimate speech activity of each speaker while it has the advantages of (i) handling unlimited number of speakers, (ii) leveraging linguistic information for speaker diarization, and (iii) simultaneously generating speaker-attributed transcriptions. Experimental results on the LibriCSS and AMI corpora show that the proposed method achieves significantly better diarization error rate than various existing speaker diarization methods when the number of speakers is unknown, and achieves a comparable performance to TS-VAD when the number of speakers is given in advance. The proposed method simultaneously generates speaker-attributed transcription with state-of-the-art accuracy.

Zhong Meng, Yashesh Gaur, Naoyuki Kanda et al.

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.

Desh Raj, Liang Lu, Zhuo Chen et al.

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.

Naoyuki Kanda, Xiong Xiao, Jian Wu et al.

Speaker-attributed automatic speech recognition (SA-ASR) is a task to recognize "who spoke what" from multi-talker recordings. An SA-ASR system usually consists of multiple modules such as speech separation, speaker diarization and ASR. On the other hand, considering the joint optimization, an end-to-end (E2E) SA-ASR model has recently been proposed with promising results on simulation data. In this paper, we present our recent study on the comparison of such modular and joint approaches towards SA-ASR on real monaural recordings. We develop state-of-the-art SA-ASR systems for both modular and joint approaches by leveraging large-scale training data, including 75 thousand hours of ASR training data and the VoxCeleb corpus for speaker representation learning. We also propose a new pipeline that performs the E2E SA-ASR model after speaker clustering. Our evaluation on the AMI meeting corpus reveals that after fine-tuning with a small real data, the joint system performs 8.9--29.9% better in accuracy compared to the best modular system while the modular system performs better before such fine-tuning. We also conduct various error analyses to show the remaining issues for the monaural SA-ASR.

Dimitrios Dimitriadis, Kenichi Kumatani, Robert Gmyr et al.

In this paper, a new learning algorithm for Federated Learning (FL) is introduced. The proposed scheme is based on a weighted gradient aggregation using two-step optimization to offer a flexible training pipeline. Herein, two different flavors of the aggregation method are presented, leading to an order of magnitude improvement in convergence speed compared to other distributed or FL training algorithms like BMUF and FedAvg. Further, the aggregation algorithm acts as a regularizer of the gradient quality. We investigate the effect of our FL algorithm in supervised and unsupervised Speech Recognition (SR) scenarios. The experimental validation is performed based on three tasks: first, the LibriSpeech task showing a speed-up of 7x and 6% word error rate reduction (WERR) compared to the baseline results. The second task is based on session adaptation providing 20% WERR over a powerful LAS model. Finally, our unsupervised pipeline is applied to the conversational SR task. The proposed FL system outperforms the baseline systems in both convergence speed and overall model performance.

Naoyuki Kanda, Guoli Ye, Yashesh Gaur et al.

This paper presents our recent effort on end-to-end speaker-attributed automatic speech recognition, which jointly performs speaker counting, speech recognition and speaker identification for monaural multi-talker audio. Firstly, we thoroughly update the model architecture that was previously designed based on a long short-term memory (LSTM)-based attention encoder decoder by applying transformer architectures. Secondly, we propose a speaker deduplication mechanism to reduce speaker identification errors in highly overlapped regions. Experimental results on the LibriSpeechMix dataset shows that the transformer-based architecture is especially good at counting the speakers and that the proposed model reduces the speaker-attributed word error rate by 47% over the LSTM-based baseline. Furthermore, for the LibriCSS dataset, which consists of real recordings of overlapped speech, the proposed model achieves concatenated minimum-permutation word error rates of 11.9% and 16.3% with and without target speaker profiles, respectively, both of which are the state-of-the-art results for LibriCSS with the monaural setting.

Naoyuki Kanda, Guoli Ye, Yu Wu et al.

Transcribing meetings containing overlapped speech with only a single distant microphone (SDM) has been one of the most challenging problems for automatic speech recognition (ASR). While various approaches have been proposed, all previous studies on the monaural overlapped speech recognition problem were based on either simulation data or small-scale real data. In this paper, we extensively investigate a two-step approach where we first pre-train a serialized output training (SOT)-based multi-talker ASR by using large-scale simulation data and then fine-tune the model with a small amount of real meeting data. Experiments are conducted by utilizing 75 thousand (K) hours of our internal single-talker recording to simulate a total of 900K hours of multi-talker audio segments for supervised pre-training. With fine-tuning on the 70 hours of the AMI-SDM training data, our SOT ASR model achieves a word error rate (WER) of 21.2% for the AMI-SDM evaluation set while automatically counting speakers in each test segment. This result is not only significantly better than the previous state-of-the-art WER of 36.4% with oracle utterance boundary information but also better than a result by a similarly fine-tuned single-talker ASR model applied to beamformed audio.

Zhong Meng, Naoyuki Kanda, Yashesh Gaur et al.

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. 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.

Xuankai Chang, Naoyuki Kanda, Yashesh Gaur et al.

An end-to-end (E2E) speaker-attributed automatic speech recognition (SA-ASR) model was proposed recently to jointly perform speaker counting, speech recognition and speaker identification. The model achieved a low speaker-attributed word error rate (SA-WER) for monaural overlapped speech comprising an unknown number of speakers. However, the E2E modeling approach is susceptible to the mismatch between the training and testing conditions. It has yet to be investigated whether the E2E SA-ASR model works well for recordings that are much longer than samples seen during training. In this work, we first apply a known decoding technique that was developed to perform single-speaker ASR for long-form audio to our E2E SA-ASR task. Then, we propose a novel method using a sequence-to-sequence model, called hypothesis stitcher. The model takes multiple hypotheses obtained from short audio segments that are extracted from the original long-form input, and it then outputs a fused single hypothesis. We propose several architectural variations of the hypothesis stitcher model and compare them with the conventional decoding methods. Experiments using LibriSpeech and LibriCSS corpora show that the proposed method significantly improves SA-WER especially for long-form multi-talker recordings.

Shahram Ghorbani, Yashesh Gaur, Yu Shi et al.

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 attends over them. We show that this architecture is better than fusing modalities at the signal level. Additionally, we also explore leveraging 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 recovering the masked words by 59% relative.

Xiaofei Wang, Naoyuki Kanda, Yashesh Gaur et al.

Joint optimization of multi-channel front-end and automatic speech recognition (ASR) has attracted much interest. While promising results have been reported for various tasks, past studies on its meeting transcription application were limited to small scale experiments. It is still unclear whether such a joint framework can be beneficial for a more practical setup where a massive amount of single channel training data can be leveraged for building a strong ASR back-end. In this work, we present our investigation on the joint modeling of a mask-based beamformer and Attention-Encoder-Decoder-based ASR in the setting where we have 75k hours of single-channel data and a relatively small amount of real multi-channel data for model training. We explore effective training procedures, including a comparison of simulated and real multi-channel training data. To guide the recognition towards a target speaker and deal with overlapped speech, we also explore various combinations of bias information, such as direction of arrivals and speaker profiles. We propose an effective location bias integration method called deep concatenation for the beamformer network. In our evaluation on various meeting recordings, we show that the proposed framework achieves a substantial word error rate reduction.

Naoyuki Kanda, Zhong Meng, Liang Lu et al.

Recently, an end-to-end speaker-attributed automatic speech recognition (E2E SA-ASR) model was proposed as a joint model of speaker counting, speech recognition and speaker identification for monaural overlapped speech. In the previous study, the model parameters were trained based on the speaker-attributed maximum mutual information (SA-MMI) criterion, with which the joint posterior probability for multi-talker transcription and speaker identification are maximized over training data. Although SA-MMI training showed promising results for overlapped speech consisting of various numbers of speakers, the training criterion was not directly linked to the final evaluation metric, i.e., speaker-attributed word error rate (SA-WER). In this paper, we propose a speaker-attributed minimum Bayes risk (SA-MBR) training method where the parameters are trained to directly minimize the expected SA-WER over the training data. Experiments using the LibriSpeech corpus show that the proposed SA-MBR training reduces the SA-WER by 9.0 % relative compared with the SA-MMI-trained model.

Zhong Meng, Sarangarajan Parthasarathy, Eric Sun et al.

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.

Naoyuki Kanda, Xuankai Chang, Yashesh Gaur et al.

Recently, an end-to-end (E2E) speaker-attributed automatic speech recognition (SA-ASR) model was proposed as a joint model of speaker counting, speech recognition and speaker identification for monaural overlapped speech. It showed promising results for simulated speech mixtures consisting of various numbers of speakers. However, the model required prior knowledge of speaker profiles to perform speaker identification, which significantly limited the application of the model. In this paper, we extend the prior work by addressing the case where no speaker profile is available. Specifically, we perform speaker counting and clustering by using the internal speaker representations of the E2E SA-ASR model to diarize the utterances of the speakers whose profiles are missing from the speaker inventory. We also propose a simple modification to the reference labels of the E2E SA-ASR training which helps handle continuous multi-talker recordings well. We conduct a comprehensive investigation of the original E2E SA-ASR and the proposed method on the monaural LibriCSS dataset. Compared to the original E2E SA-ASR with relevant speaker profiles, the proposed method achieves a close performance without any prior speaker knowledge. We also show that the source-target attention in the E2E SA-ASR model provides information about the start and end times of the hypotheses.

Dimitrios Dimitriadis, Kenichi Kumatani, Robert Gmyr et al.

In this paper, a Federated Learning (FL) simulation platform is introduced. The target scenario is Acoustic Model training based on this platform. To our knowledge, this is the first attempt to apply FL techniques to Speech Recognition tasks due to the inherent complexity. The proposed FL platform can support different tasks based on the adopted modular design. As part of the platform, a novel hierarchical optimization scheme and two gradient aggregation methods are proposed, leading to almost an order of magnitude improvement in training convergence speed compared to other distributed or FL training algorithms like BMUF and FedAvg. The hierarchical optimization offers additional flexibility in the training pipeline besides the enhanced convergence speed. On top of the hierarchical optimization, a dynamic gradient aggregation algorithm is proposed, based on a data-driven weight inference. This aggregation algorithm acts as a regularizer of the gradient quality. Finally, an unsupervised training pipeline tailored to FL is presented as a separate training scenario. The experimental validation of the proposed system is based on two tasks: first, the LibriSpeech task showing a speed-up of 7x and 6% Word Error Rate reduction (WERR) compared to the baseline results. The second task is based on session adaptation providing an improvement of 20% WERR over a competitive production-ready LAS model. The proposed Federated Learning system is shown to outperform the golden standard of distributed training in both convergence speed and overall model performance.

Naoyuki Kanda, Yashesh Gaur, Xiaofei Wang et al.

We propose an end-to-end speaker-attributed automatic speech recognition model that unifies speaker counting, speech recognition, and speaker identification on monaural overlapped speech. Our model is built on serialized output training (SOT) with attention-based encoder-decoder, a recently proposed method for recognizing overlapped speech comprising an arbitrary number of speakers. We extend SOT by introducing a speaker inventory as an auxiliary input to produce speaker labels as well as multi-speaker transcriptions. All model parameters are optimized by speaker-attributed maximum mutual information criterion, which represents a joint probability for overlapped speech recognition and speaker identification. Experiments on LibriSpeech corpus show that our proposed method achieves significantly better speaker-attributed word error rate than the baseline that separately performs overlapped speech recognition and speaker identification.

Jinyu Li, Yu Wu, Yashesh Gaur et al.

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 attention-based 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.

Hirofumi Inaguma, Yashesh Gaur, Liang Lu et al.

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 minimize 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.

Naoyuki Kanda, Yashesh Gaur, Xiaofei Wang et al.

This paper proposes serialized output training (SOT), a novel framework for multi-speaker overlapped speech recognition based on an attention-based encoder-decoder approach. Instead of having multiple output layers as with the permutation invariant training (PIT), SOT uses a model with only one output layer that generates the transcriptions of multiple speakers one after another. The attention and decoder modules take care of producing multiple transcriptions from overlapped speech. SOT has two advantages over PIT: (1) no limitation in the maximum number of speakers, and (2) an ability to model the dependencies among outputs for different speakers. We also propose a simple trick that allows SOT to be executed in $O(S)$, where $S$ is the number of the speakers in the training sample, by using the start times of the constituent source utterances. Experimental results on LibriSpeech corpus show that the SOT models can transcribe overlapped speech with variable numbers of speakers significantly better than PIT-based models. We also show that the SOT models can accurately count the number of speakers in the input audio.

Zhong Meng, Jinyu Li, Yashesh Gaur et al.

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 achieve 6.3% and 10.3% relative word error rate improvement over a strong E2E model trained with the same amount of far-field data.

Zhong Meng, Yashesh Gaur, Jinyu Li et al.

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.

Zhong Meng, Yashesh Gaur, Jinyu Li et al.

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.

Anuroop Sriram, Heewoo Jun, Yashesh Gaur et al.

This paper describes a general, scalable, end-to-end framework that uses the generative adversarial network (GAN) objective to enable robust speech recognition. Encoders trained with the proposed approach enjoy improved invariance by learning to map noisy audio to the same embedding space as that of clean audio. Unlike previous methods, the new framework does not rely on domain expertise or simplifying assumptions as are often needed in signal processing, and directly encourages robustness in a data-driven way. We show the new approach improves simulated far-field speech recognition of vanilla sequence-to-sequence models without specialized front-ends or preprocessing.

Eric Battenberg, Jitong Chen, Rewon Child et al.

In this work, we perform an empirical comparison among the CTC, RNN-Transducer, and attention-based Seq2Seq models for end-to-end speech recognition. We show that, without any language model, Seq2Seq and RNN-Transducer models both outperform the best reported CTC models with a language model, on the popular Hub5'00 benchmark. On our internal diverse dataset, these trends continue - RNNTransducer models rescored with a language model after beam search outperform our best CTC models. These results simplify the speech recognition pipeline so that decoding can now be expressed purely as neural network operations. We also study how the choice of encoder architecture affects the performance of the three models - when all encoder layers are forward only, and when encoders downsample the input representation aggressively.

Eric Battenberg, Rewon Child, Adam Coates et al.

Replacing hand-engineered pipelines with end-to-end deep learning systems has enabled strong results in applications like speech and object recognition. However, the causality and latency constraints of production systems put end-to-end speech models back into the underfitting regime and expose biases in the model that we show cannot be overcome by "scaling up", i.e., training bigger models on more data. In this work we systematically identify and address sources of bias, reducing error rates by up to 20% while remaining practical for deployment. We achieve this by utilizing improved neural architectures for streaming inference, solving optimization issues, and employing strategies that increase audio and label modelling versatility.