Rohit Prabhavalkar

Yu Zhang, Wei Han, James Qin et al. · meta-ai

We introduce the Universal Speech Model (USM), a single large model that performs automatic speech recognition (ASR) across 100+ languages. This is achieved by pre-training the encoder of the model on a large unlabeled multilingual dataset of 12 million (M) hours spanning over 300 languages, and fine-tuning on a smaller labeled dataset. We use multilingual pre-training with random-projection quantization and speech-text modality matching to achieve state-of-the-art performance on downstream multilingual ASR and speech-to-text translation tasks. We also demonstrate that despite using a labeled training set 1/7-th the size of that used for the Whisper model, our model exhibits comparable or better performance on both in-domain and out-of-domain speech recognition tasks across many languages.

Chao-Han Huck Yang, Bo Li, Yu Zhang et al. · nvidia

In this work, we propose a new parameter-efficient learning framework based on neural model reprogramming for cross-lingual speech recognition, which can \textbf{re-purpose} well-trained English automatic speech recognition (ASR) models to recognize the other languages. We design different auxiliary neural architectures focusing on learnable pre-trained feature enhancement that, for the first time, empowers model reprogramming on ASR. Specifically, we investigate how to select trainable components (i.e., encoder) of a conformer-based RNN-Transducer, as a frozen pre-trained backbone. Experiments on a seven-language multilingual LibriSpeech speech (MLS) task show that model reprogramming only requires 4.2% (11M out of 270M) to 6.8% (45M out of 660M) of its original trainable parameters from a full ASR model to perform competitive results in a range of 11.9% to 8.1% WER averaged across different languages. In addition, we discover different setups to make large-scale pre-trained ASR succeed in both monolingual and multilingual speech recognition. Our methods outperform existing ASR tuning architectures and their extension with self-supervised losses (e.g., w2v-bert) in terms of lower WER and better training efficiency.

Rohit Prabhavalkar, Takaaki Hori, Tara N. Sainath et al.

In the last decade of automatic speech recognition (ASR) research, the introduction of deep learning brought considerable reductions in word error rate of more than 50% relative, compared to modeling without deep learning. In the wake of this transition, a number of all-neural ASR architectures were introduced. These so-called end-to-end (E2E) models provide highly integrated, completely neural ASR models, which rely strongly on general machine learning knowledge, learn more consistently from data, while depending less on ASR domain-specific experience. The success and enthusiastic adoption of deep learning accompanied by more generic model architectures lead to E2E models now becoming the prominent ASR approach. The goal of this survey is to provide a taxonomy of E2E ASR models and corresponding improvements, and to discuss their properties and their relation to the classical hidden Markov model (HMM) based ASR architecture. All relevant aspects of E2E ASR are covered in this work: modeling, training, decoding, and external language model integration, accompanied by discussions of performance and deployment opportunities, as well as an outlook into potential future developments.

Tara N. Sainath, Rohit Prabhavalkar, Ankur Bapna et al.

We present JOIST, an algorithm to train a streaming, cascaded, encoder end-to-end (E2E) model with both speech-text paired inputs, and text-only unpaired inputs. Unlike previous works, we explore joint training with both modalities, rather than pre-training and fine-tuning. In addition, we explore JOIST using a streaming E2E model with an order of magnitude more data, which are also novelties compared to previous works. Through a series of ablation studies, we explore different types of text modeling, including how to model the length of the text sequence and the appropriate text sub-word unit representation. We find that best text representation for JOIST improves WER across a variety of search and rare-word test sets by 4-14% relative, compared to a model not trained with text. In addition, we quantitatively show that JOIST maintains streaming capabilities, which is important for good user-level experience.

W. Ronny Huang, Shuo-yiin Chang, David Rybach et al.

Improving the performance of end-to-end ASR models on long utterances ranging from minutes to hours in length is an ongoing challenge in speech recognition. A common solution is to segment the audio in advance using a separate voice activity detector (VAD) that decides segment boundary locations based purely on acoustic speech/non-speech information. VAD segmenters, however, may be sub-optimal for real-world speech where, e.g., a complete sentence that should be taken as a whole may contain hesitations in the middle ("set an alarm for... 5 o'clock"). We propose to replace the VAD with an end-to-end ASR model capable of predicting segment boundaries in a streaming fashion, allowing the segmentation decision to be conditioned not only on better acoustic features but also on semantic features from the decoded text with negligible extra computation. In experiments on real world long-form audio (YouTube) with lengths of up to 30 minutes, we demonstrate 8.5% relative WER improvement and 250 ms reduction in median end-of-segment latency compared to the VAD segmenter baseline on a state-of-the-art Conformer RNN-T model.

Shaojin Ding, Weiran Wang, Ding Zhao et al.

In this paper, we propose a dynamic cascaded encoder Automatic Speech Recognition (ASR) model, which unifies models for different deployment scenarios. Moreover, the model can significantly reduce model size and power consumption without loss of quality. Namely, with the dynamic cascaded encoder model, we explore three techniques to maximally boost the performance of each model size: 1) Use separate decoders for each sub-model while sharing the encoders; 2) Use funnel-pooling to improve the encoder efficiency; 3) Balance the size of causal and non-causal encoders to improve quality and fit deployment constraints. Overall, the proposed large-medium model has 30% smaller size and reduces power consumption by 33%, compared to the baseline cascaded encoder model. The triple-size model that unifies the large, medium, and small models achieves 37% total size reduction with minimal quality loss, while substantially reducing the engineering efforts of having separate models.

Zhong Meng, Weiran Wang, Rohit Prabhavalkar et al.

We propose JEIT, a joint end-to-end (E2E) model and internal language model (ILM) training method to inject large-scale unpaired text into ILM during E2E training which improves rare-word speech recognition. With JEIT, the E2E model computes an E2E loss on audio-transcript pairs while its ILM estimates a cross-entropy loss on unpaired text. The E2E model is trained to minimize a weighted sum of E2E and ILM losses. During JEIT, ILM absorbs knowledge from unpaired text while the E2E training serves as regularization. Unlike ILM adaptation methods, JEIT does not require a separate adaptation step and avoids the need for Kullback-Leibler divergence regularization of ILM. We also show that modular hybrid autoregressive transducer (MHAT) performs better than HAT in the JEIT framework, and is much more robust than HAT during ILM adaptation. To push the limit of unpaired text injection, we further propose a combined JEIT and JOIST training (CJJT) that benefits from modality matching, encoder text injection and ILM training. Both JEIT and CJJT can foster a more effective LM fusion. With 100B unpaired sentences, JEIT/CJJT improves rare-word recognition accuracy by up to 16.4% over a model trained without unpaired text.

Weiran Wang, Tongzhou Chen, Tara N. Sainath et al.

Language models (LMs) significantly improve the recognition accuracy of end-to-end (E2E) models on words rarely seen during training, when used in either the shallow fusion or the rescoring setups. In this work, we introduce LMs in the learning of hybrid autoregressive transducer (HAT) models in the discriminative training framework, to mitigate the training versus inference gap regarding the use of LMs. For the shallow fusion setup, we use LMs during both hypotheses generation and loss computation, and the LM-aware MWER-trained model achieves 10\% relative improvement over the model trained with standard MWER on voice search test sets containing rare words. For the rescoring setup, we learn a small neural module to generate per-token fusion weights in a data-dependent manner. This model achieves the same rescoring WER as regular MWER-trained model, but without the need for sweeping fusion weights.

Zhong Meng, Tongzhou Chen, Rohit Prabhavalkar et al.

Text-only adaptation of a transducer model remains challenging for end-to-end speech recognition since the transducer has no clearly separated acoustic model (AM), language model (LM) or blank model. In this work, we propose a modular hybrid autoregressive transducer (MHAT) that has structurally separated label and blank decoders to predict label and blank distributions, respectively, along with a shared acoustic encoder. The encoder and label decoder outputs are directly projected to AM and internal LM scores and then added to compute label posteriors. We train MHAT with an internal LM loss and a HAT loss to ensure that its internal LM becomes a standalone neural LM that can be effectively adapted to text. Moreover, text adaptation of MHAT fosters a much better LM fusion than internal LM subtraction-based methods. On Google's large-scale production data, a multi-domain MHAT adapted with 100B sentences achieves relative WER reductions of up to 12.4% without LM fusion and 21.5% with LM fusion from 400K-hour trained HAT.

Steven M. Hernandez, Ding Zhao, Shaojin Ding et al.

Continued improvements in machine learning techniques offer exciting new opportunities through the use of larger models and larger training datasets. However, there is a growing need to offer these new capabilities on-board low-powered devices such as smartphones, wearables and other embedded environments where only low memory is available. Towards this, we consider methods to reduce the model size of Conformer-based speech recognition models which typically require models with greater than 100M parameters down to just $5$M parameters while minimizing impact on model quality. Such a model allows us to achieve always-on ambient speech recognition on edge devices with low-memory neural processors. We propose model weight reuse at different levels within our model architecture: (i) repeating full conformer block layers, (ii) sharing specific conformer modules across layers, (iii) sharing sub-components per conformer module, and (iv) sharing decomposed sub-component weights after low-rank decomposition. By sharing weights at different levels of our model, we can retain the full model in-memory while increasing the number of virtual transformations applied to the input. Through a series of ablation studies and evaluations, we find that with weight sharing and a low-rank architecture, we can achieve a WER of 2.84 and 2.94 for Librispeech dev-clean and test-clean respectively with a $5$M parameter model.

Ke Hu, Tara N. Sainath, Yanzhang He et al.

Text-only and semi-supervised training based on audio-only data has gained popularity recently due to the wide availability of unlabeled text and speech data. In this work, we propose incorporating text-only and semi-supervised training into an attention-based deliberation model. By incorporating text-only data in training a bidirectional encoder representation from transformer (BERT) for the deliberation text encoder, and large-scale text-to-speech and audio-only utterances using joint acoustic and text decoder (JATD) and semi-supervised training, we achieved 4%-12% WER reduction for various tasks compared to the baseline deliberation. Compared to a state-of-the-art language model (LM) rescoring method, the deliberation model reduces the Google Voice Search WER by 11% relative. We show that the deliberation model also achieves a positive human side-by-side evaluation compared to the state-of-the-art LM rescorer with reasonable endpointer latencies.

Zih-Ching Chen, Chao-Han Huck Yang, Bo Li et al.

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

Cal Peyser, Zhong Meng, Ke Hu et al.

The last year has seen astonishing progress in text-prompted image generation premised on the idea of a cross-modal representation space in which the text and image domains are represented jointly. In ASR, this idea has found application as joint speech-text encoders that can scale to the capacities of very large parameter models by being trained on both unpaired speech and text. While these methods show promise, they have required special treatment of the sequence-length mismatch inherent in speech and text, either by up-sampling heuristics or an explicit alignment model. In this work, we offer evidence that joint speech-text encoders naturally achieve consistent representations across modalities by disregarding sequence length, and argue that consistency losses could forgive length differences and simply assume the best alignment. We show that such a loss improves downstream WER in both a large-parameter monolingual and multilingual system.

W. Ronny Huang, Shuo-Yiin Chang, Tara N. Sainath et al.

We explore unifying a neural segmenter with two-pass cascaded encoder ASR into a single model. A key challenge is allowing the segmenter (which runs in real-time, synchronously with the decoder) to finalize the 2nd pass (which runs 900 ms behind real-time) without introducing user-perceived latency or deletion errors during inference. We propose a design where the neural segmenter is integrated with the causal 1st pass decoder to emit a end-of-segment (EOS) signal in real-time. The EOS signal is then used to finalize the non-causal 2nd pass. We experiment with different ways to finalize the 2nd pass, and find that a novel dummy frame injection strategy allows for simultaneous high quality 2nd pass results and low finalization latency. On a real-world long-form captioning task (YouTube), we achieve 2.4% relative WER and 140 ms EOS latency gains over a baseline VAD-based segmenter with the same cascaded encoder.

Cal Peyser, Michael Picheny, Kyunghyun Cho et al.

Unpaired text and audio injection have emerged as dominant methods for improving ASR performance in the absence of a large labeled corpus. However, little guidance exists on deploying these methods to improve production ASR systems that are trained on very large supervised corpora and with realistic requirements like a constrained model size and CPU budget, streaming capability, and a rich lattice for rescoring and for downstream NLU tasks. In this work, we compare three state-of-the-art semi-supervised methods encompassing both unpaired text and audio as well as several of their combinations in a controlled setting using joint training. We find that in our setting these methods offer many improvements beyond raw WER, including substantial gains in tail-word WER, decoder computation during inference, and lattice density.

Rami Botros, Rohit Prabhavalkar, Johan Schalkwyk et al.

In end-to-end (E2E) speech recognition models, a representational tight-coupling inevitably emerges between the encoder and the decoder. We build upon recent work that has begun to explore building encoders with modular encoded representations, such that encoders and decoders from different models can be stitched together in a zero-shot manner without further fine-tuning. While previous research only addresses full-context speech models, we explore the problem in a streaming setting as well. Our framework builds on top of existing encoded representations, converting them to modular features, dubbed as Lego-Features, without modifying the pre-trained model. The features remain interchangeable when the model is retrained with distinct initializations. Though sparse, we show that the Lego-Features are powerful when tested with RNN-T or LAS decoders, maintaining high-quality downstream performance. They are also rich enough to represent the first-pass prediction during two-pass deliberation. In this scenario, they outperform the N-best hypotheses, since they do not need to be supplemented with acoustic features to deliver the best results. Moreover, generating the Lego-Features does not require beam search or auto-regressive computation. Overall, they present a modular, powerful and cheap alternative to the standard encoder output, as well as the N-best hypotheses.

Weiran Wang, Zelin Wu, Diamantino Caseiro et al.

Contextual biasing refers to the problem of biasing the automatic speech recognition (ASR) systems towards rare entities that are relevant to the specific user or application scenarios. We propose algorithms for contextual biasing based on the Knuth-Morris-Pratt algorithm for pattern matching. During beam search, we boost the score of a token extension if it extends matching into a set of biasing phrases. Our method simulates the classical approaches often implemented in the weighted finite state transducer (WFST) framework, but avoids the FST language altogether, with careful considerations on memory footprint and efficiency on tensor processing units (TPUs) by vectorization. Without introducing additional model parameters, our method achieves significant word error rate (WER) reductions on biasing test sets by itself, and yields further performance gain when combined with a model-based biasing method.

Cal Peyser, Ronny Huang, Tara Sainath et al.

Dual learning is a paradigm for semi-supervised machine learning that seeks to leverage unsupervised data by solving two opposite tasks at once. In this scheme, each model is used to generate pseudo-labels for unlabeled examples that are used to train the other model. Dual learning has seen some use in speech processing by pairing ASR and TTS as dual tasks. However, these results mostly address only the case of using unpaired examples to compensate for very small supervised datasets, and mostly on large, non-streaming models. Dual learning has not yet been proven effective for using unsupervised data to improve realistic on-device streaming models that are already trained on large supervised corpora. We provide this missing piece though an analysis of an on-device-sized streaming conformer trained on the entirety of Librispeech, showing relative WER improvements of 10.7%/5.2% without an LM and 11.7%/16.4% with an LM.

Lillian Zhou, Yuxin Ding, Mingqing Chen et al.

Automatic speech recognition (ASR) models are typically trained on large datasets of transcribed speech. As language evolves and new terms come into use, these models can become outdated and stale. In the context of models trained on the server but deployed on edge devices, errors may result from the mismatch between server training data and actual on-device usage. In this work, we seek to continually learn from on-device user corrections through Federated Learning (FL) to address this issue. We explore techniques to target fresh terms that the model has not previously encountered, learn long-tail words, and mitigate catastrophic forgetting. In experimental evaluations, we find that the proposed techniques improve model recognition of fresh terms, while preserving quality on the overall language distribution.

Rohit Prabhavalkar, Zhong Meng, Weiran Wang et al.

The accuracy of end-to-end (E2E) automatic speech recognition (ASR) models continues to improve as they are scaled to larger sizes, with some now reaching billions of parameters. Widespread deployment and adoption of these models, however, requires computationally efficient strategies for decoding. In the present work, we study one such strategy: applying multiple frame reduction layers in the encoder to compress encoder outputs into a small number of output frames. While similar techniques have been investigated in previous work, we achieve dramatically more reduction than has previously been demonstrated through the use of multiple funnel reduction layers. Through ablations, we study the impact of various architectural choices in the encoder to identify the most effective strategies. We demonstrate that we can generate one encoder output frame for every 2.56 sec of input speech, without significantly affecting word error rate on a large-scale voice search task, while improving encoder and decoder latencies by 48% and 92% respectively, relative to a strong but computationally expensive baseline.

Zelin Wu, Gan Song, Christopher Li et al.

Contextual biasing enables speech recognizers to transcribe important phrases in the speaker's context, such as contact names, even if they are rare in, or absent from, the training data. Attention-based biasing is a leading approach which allows for full end-to-end cotraining of the recognizer and biasing system and requires no separate inference-time components. Such biasers typically consist of a context encoder; followed by a context filter which narrows down the context to apply, improving per-step inference time; and, finally, context application via cross attention. Though much work has gone into optimizing per-frame performance, the context encoder is at least as important: recognition cannot begin before context encoding ends. Here, we show the lightweight phrase selection pass can be moved before context encoding, resulting in a speedup of up to 16.1 times and enabling biasing to scale to 20K phrases with a maximum pre-decoding delay under 33ms. With the addition of phrase- and wordpiece-level cross-entropy losses, our technique also achieves up to a 37.5% relative WER reduction over the baseline without the losses and lightweight phrase selection pass.

Guru Prakash Arumugam, Shuo-yiin Chang, Tara N. Sainath et al.

ASR models often suffer from a long-form deletion problem where the model predicts sequential blanks instead of words when transcribing a lengthy audio (in the order of minutes or hours). From the perspective of a user or downstream system consuming the ASR results, this behavior can be perceived as the model "being stuck", and potentially make the product hard to use. One of the culprits for long-form deletion is training-test data mismatch, which can happen even when the model is trained on diverse and large-scale data collected from multiple application domains. In this work, we introduce a novel technique to simultaneously model different groups of speakers in the audio along with the standard transcript tokens. Speakers are grouped as primary and non-primary, which connects the application domains and significantly alleviates the long-form deletion problem. This improved model neither needs any additional training data nor incurs additional training or inference cost.

Adam Stooke, Rohit Prabhavalkar, Khe Chai Sim et al.

Modern systems for automatic speech recognition, including the RNN-Transducer and Attention-based Encoder-Decoder (AED), are designed so that the encoder is not required to alter the time-position of information from the audio sequence into the embedding; alignment to the final text output is processed during decoding. We discover that the transformer-based encoder adopted in recent years is actually capable of performing the alignment internally during the forward pass, prior to decoding. This new phenomenon enables a simpler and more efficient model, the "Aligner-Encoder". To train it, we discard the dynamic programming of RNN-T in favor of the frame-wise cross-entropy loss of AED, while the decoder employs the lighter text-only recurrence of RNN-T without learned cross-attention -- it simply scans embedding frames in order from the beginning, producing one token each until predicting the end-of-message. We conduct experiments demonstrating performance remarkably close to the state of the art, including a special inference configuration enabling long-form recognition. In a representative comparison, we measure the total inference time for our model to be 2x faster than RNN-T and 16x faster than AED. Lastly, we find that the audio-text alignment is clearly visible in the self-attention weights of a certain layer, which could be said to perform "self-transduction".

Zhong Meng, Zelin Wu, Rohit Prabhavalkar et al.

Neural contextual biasing effectively improves automatic speech recognition (ASR) for crucial phrases within a speaker's context, particularly those that are infrequent in the training data. This work proposes contextual text injection (CTI) to enhance contextual ASR. CTI leverages not only the paired speech-text data, but also a much larger corpus of unpaired text to optimize the ASR model and its biasing component. Unpaired text is converted into speech-like representations and used to guide the model's attention towards relevant bias phrases. Moreover, we introduce a contextual text-injected (CTI) minimum word error rate (MWER) training, which minimizes the expected WER caused by contextual biasing when unpaired text is injected into the model. Experiments show that CTI with 100 billion text sentences can achieve up to 43.3% relative WER reduction from a strong neural biasing model. CTI-MWER provides a further relative improvement of 23.5%.

Lun Wang, Om Thakkar, Zhong Meng et al.

Gradient clipping plays a vital role in training large-scale automatic speech recognition (ASR) models. It is typically applied to minibatch gradients to prevent gradient explosion, and to the individual sample gradients to mitigate unintended memorization. This work systematically investigates the impact of a specific granularity of gradient clipping, namely per-core clip-ping (PCC), across training a wide range of ASR models. We empirically demonstrate that PCC can effectively mitigate unintended memorization in ASR models. Surprisingly, we find that PCC positively influences ASR performance metrics, leading to improved convergence rates and reduced word error rates. To avoid tuning the additional hyperparameter introduced by PCC, we further propose a novel variant, adaptive per-core clipping (APCC), for streamlined optimization. Our findings highlight the multifaceted benefits of PCC as a strategy for robust, privacy-forward ASR model training.

Antoine Bruguier, Duc Le, Rohit Prabhavalkar et al.

We propose Neural-FST Class Language Model (NFCLM) for end-to-end speech recognition, a novel method that combines neural network language models (NNLMs) and finite state transducers (FSTs) in a mathematically consistent framework. Our method utilizes a background NNLM which models generic background text together with a collection of domain-specific entities modeled as individual FSTs. Each output token is generated by a mixture of these components; the mixture weights are estimated with a separately trained neural decider. We show that NFCLM significantly outperforms NNLM by 15.8% relative in terms of Word Error Rate. NFCLM achieves similar performance as traditional NNLM and FST shallow fusion while being less prone to overbiasing and 12 times more compact, making it more suitable for on-device usage.

Zhiyun Lu, Yanwei Pan, Thibault Doutre et al.

End-to-end models have achieved state-of-the-art results on several automatic speech recognition tasks. However, they perform poorly when evaluated on long-form data, e.g., minutes long conversational telephony audio. One reason the model fails on long-form speech is that it has only seen short utterances during training. In this paper we study the effect of training utterance length on the word error rate (WER) for RNN-transducer (RNN-T) model. We compare two widely used training objectives, log loss (or RNN-T loss) and minimum word error rate (MWER) loss. We conduct experiments on telephony datasets in four languages. Our experiments show that for both losses, the WER on long-form speech reduces substantially as the training utterance length increases. The average relative WER gain is 15.7% for log loss and 8.8% for MWER loss. When training on short utterances, MWER loss leads to a lower WER than the log loss. Such difference between the two losses diminishes when the input length increases.

Nathan Howard, Alex Park, Turaj Zakizadeh Shabestary et al.

We consider the problem of recognizing speech utterances spoken to a device which is generating a known sound waveform; for example, recognizing queries issued to a digital assistant which is generating responses to previous user inputs. Previous work has proposed building acoustic echo cancellation (AEC) models for this task that optimize speech enhancement metrics using both neural network as well as signal processing approaches. Since our goal is to recognize the input speech, we consider enhancements which improve word error rates (WERs) when the predicted speech signal is passed to an automatic speech recognition (ASR) model. First, we augment the loss function with a term that produces outputs useful to a pre-trained ASR model and show that this augmented loss function improves WER metrics. Second, we demonstrate that augmenting our training dataset of real world examples with a large synthetic dataset improves performance. Crucially, applying SpecAugment style masks to the reference channel during training aids the model in adapting from synthetic to real domains. In experimental evaluations, we find the proposed approaches improve performance, on average, by 57% over a signal processing baseline and 45% over the neural AEC model without the proposed changes.

Yuan Shangguan, Rohit Prabhavalkar, Hang Su et al.

As speech-enabled devices such as smartphones and smart speakers become increasingly ubiquitous, there is growing interest in building automatic speech recognition (ASR) systems that can run directly on-device; end-to-end (E2E) speech recognition models such as recurrent neural network transducers and their variants have recently emerged as prime candidates for this task. Apart from being accurate and compact, such systems need to decode speech with low user-perceived latency (UPL), producing words as soon as they are spoken. This work examines the impact of various techniques - model architectures, training criteria, decoding hyperparameters, and endpointer parameters - on UPL. Our analyses suggest that measures of model size (parameters, input chunk sizes), or measures of computation (e.g., FLOPS, RTF) that reflect the model's ability to process input frames are not always strongly correlated with observed UPL. Thus, conventional algorithmic latency measurements might be inadequate in accurately capturing latency observed when models are deployed on embedded devices. Instead, we find that factors affecting token emission latency, and endpointing behavior have a larger impact on UPL. We achieve the best trade-off between latency and word error rate when performing ASR jointly with endpointing, while utilizing the recently proposed alignment regularization mechanism.

Yangyang Shi, Varun Nagaraja, Chunyang Wu et al.

We propose a dynamic encoder transducer (DET) for on-device speech recognition. One DET model scales to multiple devices with different computation capacities without retraining or finetuning. To trading off accuracy and latency, DET assigns different encoders to decode different parts of an utterance. We apply and compare the layer dropout and the collaborative learning for DET training. The layer dropout method that randomly drops out encoder layers in the training phase, can do on-demand layer dropout in decoding. Collaborative learning jointly trains multiple encoders with different depths in one single model. Experiment results on Librispeech and in-house data show that DET provides a flexible accuracy and latency trade-off. Results on Librispeech show that the full-size encoder in DET relatively reduces the word error rate of the same size baseline by over 8%. The lightweight encoder in DET trained with collaborative learning reduces the model size by 25% but still gets similar WER as the full-size baseline. DET gets similar accuracy as a baseline model with better latency on a large in-house data set by assigning a lightweight encoder for the beginning part of one utterance and a full-size encoder for the rest.

David Qiu, Qiujia Li, Yanzhang He et al.

We study the problem of word-level confidence estimation in subword-based end-to-end (E2E) models for automatic speech recognition (ASR). Although prior works have proposed training auxiliary confidence models for ASR systems, they do not extend naturally to systems that operate on word-pieces (WP) as their vocabulary. In particular, ground truth WP correctness labels are needed for training confidence models, but the non-unique tokenization from word to WP causes inaccurate labels to be generated. This paper proposes and studies two confidence models of increasing complexity to solve this problem. The final model uses self-attention to directly learn word-level confidence without needing subword tokenization, and exploits full context features from multiple hypotheses to improve confidence accuracy. Experiments on Voice Search and long-tail test sets show standard metrics (e.g., NCE, AUC, RMSE) improving substantially. The proposed confidence module also enables a model selection approach to combine an on-device E2E model with a hybrid model on the server to address the rare word recognition problem for the E2E model.

Rohit Prabhavalkar, Yanzhang He, David Rybach et al.

End-to-end models that condition the output label sequence on all previously predicted labels have emerged as popular alternatives to conventional systems for automatic speech recognition (ASR). Since unique label histories correspond to distinct models states, such models are decoded using an approximate beam-search process which produces a tree of hypotheses. In this work, we study the influence of the amount of label context on the model's accuracy, and its impact on the efficiency of the decoding process. We find that we can limit the context of the recurrent neural network transducer (RNN-T) during training to just four previous word-piece labels, without degrading word error rate (WER) relative to the full-context baseline. Limiting context also provides opportunities to improve the efficiency of the beam-search process during decoding by removing redundant paths from the active beam, and instead retaining them in the final lattice. This path-merging scheme can also be applied when decoding the baseline full-context model through an approximation. Overall, we find that the proposed path-merging scheme is extremely effective allowing us to improve oracle WERs by up to 36% over the baseline, while simultaneously reducing the number of model evaluations by up to 5.3% without any degradation in WER.

Arun Narayanan, Tara N. Sainath, Ruoming Pang et al.

End-to-end (E2E) automatic speech recognition (ASR) models, by now, have shown competitive performance on several benchmarks. These models are structured to either operate in streaming or non-streaming mode. This work presents cascaded encoders for building a single E2E ASR model that can operate in both these modes simultaneously. The proposed model consists of streaming and non-streaming encoders. Input features are first processed by the streaming encoder; the non-streaming encoder operates exclusively on the output of the streaming encoder. A single decoder then learns to decode either using the output of the streaming or the non-streaming encoder. Results show that this model achieves similar word error rates (WER) as a standalone streaming model when operating in streaming mode, and obtains 10% -- 27% relative improvement when operating in non-streaming mode. Our results also show that the proposed approach outperforms existing E2E two-pass models, especially on long-form speech.

Daria Soboleva, Ondrej Skopek, Márius Šajgalík et al.

We present a novel multi-modal unspoken punctuation prediction system for the English language which combines acoustic and text features. We demonstrate for the first time, that by relying exclusively on synthetic data generated using a prosody-aware text-to-speech system, we can outperform a model trained with expensive human audio recordings on the unspoken punctuation prediction problem. Our model architecture is well suited for on-device use. This is achieved by leveraging hash-based embeddings of automatic speech recognition text output in conjunction with acoustic features as input to a quasi-recurrent neural network, keeping the model size small and latency low.

Chung-Cheng Chiu, Arun Narayanan, Wei Han et al.

In recent years, all-neural end-to-end approaches have obtained state-of-the-art results on several challenging automatic speech recognition (ASR) tasks. However, most existing works focus on building ASR models where train and test data are drawn from the same domain. This results in poor generalization characteristics on mismatched-domains: e.g., end-to-end models trained on short segments perform poorly when evaluated on longer utterances. In this work, we analyze the generalization properties of streaming and non-streaming recurrent neural network transducer (RNN-T) based end-to-end models in order to identify model components that negatively affect generalization performance. We propose two solutions: combining multiple regularization techniques during training, and using dynamic overlapping inference. On a long-form YouTube test set, when the nonstreaming RNN-T model is trained with shorter segments of data, the proposed combination improves word error rate (WER) from 22.3% to 14.8%; when the streaming RNN-T model trained on short Search queries, the proposed techniques improve WER on the YouTube set from 67.0% to 25.3%. Finally, when trained on Librispeech, we find that dynamic overlapping inference improves WER on YouTube from 99.8% to 33.0%.

Tara N. Sainath, Yanzhang He, Bo Li et al.

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

Ke Hu, Tara N. Sainath, Ruoming Pang et al.

End-to-end (E2E) models have made rapid progress in automatic speech recognition (ASR) and perform competitively relative to conventional models. To further improve the quality, a two-pass model has been proposed to rescore streamed hypotheses using the non-streaming Listen, Attend and Spell (LAS) model while maintaining a reasonable latency. The model attends to acoustics to rescore hypotheses, as opposed to a class of neural correction models that use only first-pass text hypotheses. In this work, we propose to attend to both acoustics and first-pass hypotheses using a deliberation network. A bidirectional encoder is used to extract context information from first-pass hypotheses. The proposed deliberation model achieves 12% relative WER reduction compared to LAS rescoring in Google Voice Search (VS) tasks, and 23% reduction on a proper noun test set. Compared to a large conventional model, our best model performs 21% relatively better for VS. In terms of computational complexity, the deliberation decoder has a larger size than the LAS decoder, and hence requires more computations in second-pass decoding.

Chung-Cheng Chiu, Wei Han, Yu Zhang et al.

End-to-end automatic speech recognition (ASR) models, including both attention-based models and the recurrent neural network transducer (RNN-T), have shown superior performance compared to conventional systems. However, previous studies have focused primarily on short utterances that typically last for just a few seconds or, at most, a few tens of seconds. Whether such architectures are practical on long utterances that last from minutes to hours remains an open question. In this paper, we both investigate and improve the performance of end-to-end models on long-form transcription. We first present an empirical comparison of different end-to-end models on a real world long-form task and demonstrate that the RNN-T model is much more robust than attention-based systems in this regime. We next explore two improvements to attention-based systems that significantly improve its performance: restricting the attention to be monotonic, and applying a novel decoding algorithm that breaks long utterances into shorter overlapping segments. Combining these two improvements, we show that attention-based end-to-end models can be very competitive to RNN-T on long-form speech recognition.

Arun Narayanan, Rohit Prabhavalkar, Chung-Cheng Chiu et al.

All-neural end-to-end (E2E) automatic speech recognition (ASR) systems that use a single neural network to transduce audio to word sequences have been shown to achieve state-of-the-art results on several tasks. In this work, we examine the ability of E2E models to generalize to unseen domains, where we find that models trained on short utterances fail to generalize to long-form speech. We propose two complementary solutions to address this: training on diverse acoustic data, and LSTM state manipulation to simulate long-form audio when training using short utterances. On a synthesized long-form test set, adding data diversity improves word error rate (WER) by 90% relative, while simulating long-form training improves it by 67% relative, though the combination doesn't improve over data diversity alone. On a real long-form call-center test set, adding data diversity improves WER by 40% relative. Simulating long-form training on top of data diversity improves performance by an additional 27% relative.

Tara N. Sainath, Ruoming Pang, David Rybach et al.

The requirements for many applications of state-of-the-art speech recognition systems include not only low word error rate (WER) but also low latency. Specifically, for many use-cases, the system must be able to decode utterances in a streaming fashion and faster than real-time. Recently, a streaming recurrent neural network transducer (RNN-T) end-to-end (E2E) model has shown to be a good candidate for on-device speech recognition, with improved WER and latency metrics compared to conventional on-device models [1]. However, this model still lags behind a large state-of-the-art conventional model in quality [2]. On the other hand, a non-streaming E2E Listen, Attend and Spell (LAS) model has shown comparable quality to large conventional models [3]. This work aims to bring the quality of an E2E streaming model closer to that of a conventional system by incorporating a LAS network as a second-pass component, while still abiding by latency constraints. Our proposed two-pass model achieves a 17%-22% relative reduction in WER compared to RNN-T alone and increases latency by a small fraction over RNN-T.

Ke Hu, Antoine Bruguier, Tara N. Sainath et al.

Contextual automatic speech recognition, i.e., biasing recognition towards a given context (e.g. user's playlists, or contacts), is challenging in end-to-end (E2E) models. Such models maintain a limited number of candidates during beam-search decoding, and have been found to recognize rare named entities poorly. The problem is exacerbated when biasing towards proper nouns in foreign languages, e.g., geographic location names, which are virtually unseen in training and are thus out-of-vocabulary (OOV). While grapheme or wordpiece E2E models might have a difficult time spelling OOV words, phonemes are more acoustically salient and past work has shown that E2E phoneme models can better predict such words. In this work, we propose an E2E model containing both English wordpieces and phonemes in the modeling space, and perform contextual biasing of foreign words at the phoneme level by mapping pronunciations of foreign words into similar English phonemes. In experimental evaluations, we find that the proposed approach performs 16% better than a grapheme-only biasing model, and 8% better than a wordpiece-only biasing model on a foreign place name recognition task, with only slight degradation on regular English tasks.

Jonathan Shen, Patrick Nguyen, Yonghui Wu et al.

Lingvo is a Tensorflow framework offering a complete solution for collaborative deep learning research, with a particular focus towards sequence-to-sequence models. Lingvo models are composed of modular building blocks that are flexible and easily extensible, and experiment configurations are centralized and highly customizable. Distributed training and quantized inference are supported directly within the framework, and it contains existing implementations of a large number of utilities, helper functions, and the newest research ideas. Lingvo has been used in collaboration by dozens of researchers in more than 20 papers over the last two years. This document outlines the underlying design of Lingvo and serves as an introduction to the various pieces of the framework, while also offering examples of advanced features that showcase the capabilities of the framework.

Kazuki Irie, Rohit Prabhavalkar, Anjuli Kannan et al.

In conventional speech recognition, phoneme-based models outperform grapheme-based models for non-phonetic languages such as English. The performance gap between the two typically reduces as the amount of training data is increased. In this work, we examine the impact of the choice of modeling unit for attention-based encoder-decoder models. We conduct experiments on the LibriSpeech 100hr, 460hr, and 960hr tasks, using various target units (phoneme, grapheme, and word-piece); across all tasks, we find that grapheme or word-piece models consistently outperform phoneme-based models, even though they are evaluated without a lexicon or an external language model. We also investigate model complementarity: we find that we can improve WERs by up to 9% relative by rescoring N-best lists generated from a strong word-piece based baseline with either the phoneme or the grapheme model. Rescoring an N-best list generated by the phonemic system, however, provides limited improvements. Further analysis shows that the word-piece-based models produce more diverse N-best hypotheses, and thus lower oracle WERs, than phonemic models.

Yanzhang He, Tara N. Sainath, Rohit Prabhavalkar et al.

End-to-end (E2E) models, which directly predict output character sequences given input speech, are good candidates for on-device speech recognition. E2E models, however, present numerous challenges: In order to be truly useful, such models must decode speech utterances in a streaming fashion, in real time; they must be robust to the long tail of use cases; they must be able to leverage user-specific context (e.g., contact lists); and above all, they must be extremely accurate. In this work, we describe our efforts at building an E2E speech recognizer using a recurrent neural network transducer. In experimental evaluations, we find that the proposed approach can outperform a conventional CTC-based model in terms of both latency and accuracy in a number of evaluation categories.

Parisa Haghani, Arun Narayanan, Michiel Bacchiani et al.

Conventional spoken language understanding systems consist of two main components: an automatic speech recognition module that converts audio to a transcript, and a natural language understanding module that transforms the resulting text (or top N hypotheses) into a set of domains, intents, and arguments. These modules are typically optimized independently. In this paper, we formulate audio to semantic understanding as a sequence-to-sequence problem [1]. We propose and compare various encoder-decoder based approaches that optimize both modules jointly, in an end-to-end manner. Evaluations on a real-world task show that 1) having an intermediate text representation is crucial for the quality of the predicted semantics, especially the intent arguments and 2) jointly optimizing the full system improves overall accuracy of prediction. Compared to independently trained models, our best jointly trained model achieves similar domain and intent prediction F1 scores, but improves argument word error rate by 18% relative.

Golan Pundak, Tara N. Sainath, Rohit Prabhavalkar et al.

In automatic speech recognition (ASR) what a user says depends on the particular context she is in. Typically, this context is represented as a set of word n-grams. In this work, we present a novel, all-neural, end-to-end (E2E) ASR sys- tem that utilizes such context. Our approach, which we re- fer to as Contextual Listen, Attend and Spell (CLAS) jointly- optimizes the ASR components along with embeddings of the context n-grams. During inference, the CLAS system can be presented with context phrases which might contain out-of- vocabulary (OOV) terms not seen during training. We com- pare our proposed system to a more traditional contextualiza- tion approach, which performs shallow-fusion between inde- pendently trained LAS and contextual n-gram models during beam search. Across a number of tasks, we find that the pro- posed CLAS system outperforms the baseline method by as much as 68% relative WER, indicating the advantage of joint optimization over individually trained components. Index Terms: speech recognition, sequence-to-sequence models, listen attend and spell, LAS, attention, embedded speech recognition.

Kanishka Rao, Haşim Sak, Rohit Prabhavalkar

We investigate training end-to-end speech recognition models with the recurrent neural network transducer (RNN-T): a streaming, all-neural, sequence-to-sequence architecture which jointly learns acoustic and language model components from transcribed acoustic data. We explore various model architectures and demonstrate how the model can be improved further if additional text or pronunciation data are available. The model consists of an `encoder', which is initialized from a connectionist temporal classification-based (CTC) acoustic model, and a `decoder' which is partially initialized from a recurrent neural network language model trained on text data alone. The entire neural network is trained with the RNN-T loss and directly outputs the recognized transcript as a sequence of graphemes, thus performing end-to-end speech recognition. We find that performance can be improved further through the use of sub-word units (`wordpieces') which capture longer context and significantly reduce substitution errors. The best RNN-T system, a twelve-layer LSTM encoder with a two-layer LSTM decoder trained with 30,000 wordpieces as output targets achieves a word error rate of 8.5\% on voice-search and 5.2\% on voice-dictation tasks and is comparable to a state-of-the-art baseline at 8.3\% on voice-search and 5.4\% voice-dictation.

Anjuli Kannan, Yonghui Wu, Patrick Nguyen et al.

Attention-based sequence-to-sequence models for automatic speech recognition jointly train an acoustic model, language model, and alignment mechanism. Thus, the language model component is only trained on transcribed audio-text pairs. This leads to the use of shallow fusion with an external language model at inference time. Shallow fusion refers to log-linear interpolation with a separately trained language model at each step of the beam search. In this work, we investigate the behavior of shallow fusion across a range of conditions: different types of language models, different decoding units, and different tasks. On Google Voice Search, we demonstrate that the use of shallow fusion with a neural LM with wordpieces yields a 9.1% relative word error rate reduction (WERR) over our competitive attention-based sequence-to-sequence model, obviating the need for second-pass rescoring.

Tara N. Sainath, Rohit Prabhavalkar, Shankar Kumar et al.

For decades, context-dependent phonemes have been the dominant sub-word unit for conventional acoustic modeling systems. This status quo has begun to be challenged recently by end-to-end models which seek to combine acoustic, pronunciation, and language model components into a single neural network. Such systems, which typically predict graphemes or words, simplify the recognition process since they remove the need for a separate expert-curated pronunciation lexicon to map from phoneme-based units to words. However, there has been little previous work comparing phoneme-based versus grapheme-based sub-word units in the end-to-end modeling framework, to determine whether the gains from such approaches are primarily due to the new probabilistic model, or from the joint learning of the various components with grapheme-based units. In this work, we conduct detailed experiments which are aimed at quantifying the value of phoneme-based pronunciation lexica in the context of end-to-end models. We examine phoneme-based end-to-end models, which are contrasted against grapheme-based ones on a large vocabulary English Voice-search task, where we find that graphemes do indeed outperform phonemes. We also compare grapheme and phoneme-based approaches on a multi-dialect English task, which once again confirm the superiority of graphemes, greatly simplifying the system for recognizing multiple dialects.

Rohit Prabhavalkar, Tara N. Sainath, Yonghui Wu et al.

Sequence-to-sequence models, such as attention-based models in automatic speech recognition (ASR), are typically trained to optimize the cross-entropy criterion which corresponds to improving the log-likelihood of the data. However, system performance is usually measured in terms of word error rate (WER), not log-likelihood. Traditional ASR systems benefit from discriminative sequence training which optimizes criteria such as the state-level minimum Bayes risk (sMBR) which are more closely related to WER. In the present work, we explore techniques to train attention-based models to directly minimize expected word error rate. We consider two loss functions which approximate the expected number of word errors: either by sampling from the model, or by using N-best lists of decoded hypotheses, which we find to be more effective than the sampling-based method. In experimental evaluations, we find that the proposed training procedure improves performance by up to 8.2% relative to the baseline system. This allows us to train grapheme-based, uni-directional attention-based models which match the performance of a traditional, state-of-the-art, discriminative sequence-trained system on a mobile voice-search task.