Jump to Content
Yanzhang He

Yanzhang He

Authored Publications
Google Publications
Other Publications
Sort By
  • Title
  • Title, desc
  • Year
  • Year, desc
    Preview abstract Text-only and semi-supervised training based on audio-only data has gained popularity recently due to the wide availability of unlabeled text or speech data. In this work, we propose text-only and semi-supervised training for attention-decoder based deliberation. By incorporating text-only data in training a bidirectional encoder representation from transformer (BERT) for the deliberation text encoder, joint acoustic and text decoder (JATD) training, and semi-supervised training based on a conventional model as a teacher, we achieved up to 11.7% WER reduction compared to the baseline deliberation. Compared to a state-of-the-art language model (LM) rescoring method, the deliberation model reduces the WER by 8% relative for Google Voice Search with reasonable endpointing latencies. We show that the deliberation has achieved a positive human side-by-side evaluation compared to LM rescoring. View details
    Preview abstract Automatic speech recognition (ASR) systems typically rely on an external endpointer (EP) model to identify speech boundaries. This EP model strongly affects latency, but is subject to computational constraints, which limits prediction accuracy. We propose a method to jointly train the ASR and EP tasks in a single end-to-end (E2E) multitask model, improving EP quality by optionally leveraging information from the ASR audio encoder. We introduce a "switch" connection, which trains the EP to consume either the audio frames directly or low-level latent representations from the ASR model. This allows flexibility during inference to produce a low-cost prediction or a higher quality prediction if ASR computation is ongoing. We present results on a voice search test set showing that, compared to separate single-task models, this approach reduces median endpoint latency by 130ms (33.3% reduction), and 90th percentile latency by 160ms (22.2% reduction), without regressing word-error rate. For continuous recognition, WER improves by 10.6% (relative). View details
    Preview abstract Self- and Semi-supervised learning methods have been actively investigated to reduce labeled training data or enhance the model performance. However, the approach mostly focus on in-domain performance for public datasets. In this study, we utilize the combination of self- and semi-supervised learning methods to solve unseen domain adaptation problem in a large-scale production setting for online ASR model. This approach demonstrates that using the source domain data with a small fraction of the target domain data (3%) can recover the performance gap compared to a full data baseline: relative 13.5% WER improvement for target domain data. View details
    Preview abstract VoiceFilter-Lite is a speaker-conditioned voice separation model that plays a crucial role in improving speech recognition and speaker verification by suppressing overlapping speech from the non-target speaker. One limitation of VoiceFilter-Lite, and other speaker-conditioned speech models in general, is that these models are usually limited to a single target speaker. This is undesirable as most smart home devices now support multiple enrolled users. In order to extend the benefits of personalization to multiple users, we previously developed an attention-based speaker selection mechanism and applied it to VoiceFilter-Lite. However, the original multi-user VoiceFilter-Lite model suffers from significant performance degradation compared with single-user models. In this paper, we devised a series of experiments to improve the multi-user VoiceFilter-Lite model. By incorporating dual learning rates and using feature-wise linear modulation (FiLM) to condition the model with the attended embedding, we successfully closed the performance gap between multi-user and single-user VoiceFilter-Lite models on single-speaker evaluations. At the same time, the new model can also be easily extended to support any number of users, and significantly outperforms our previously published model on multi-speaker evaluations. View details
    Preview abstract Streaming automatic speech recognition (ASR) aims to output each hypothesized word as quickly and accurately as possible. However, reducing latency while retaining accuracy is highly challenging. Existing approaches including Early and Late Penalties~\cite{li2020towards} and Constrained Alignment~\cite{sainath2020emitting} penalize emission delay by manipulating per-token or per-frame RNN-T output logits. While being successful in reducing latency, these approaches lead to significant accuracy degradation. In this work, we propose a sequence-level emission regularization technique, named FastEmit, that applies emission latency regularization directly on the transducer forward-backward probabilities. We demonstrate that FastEmit is more suitable to the sequence-level transducer~\cite{Graves12} training objective for streaming ASR networks. We apply FastEmit on various end-to-end (E2E) ASR networks including RNN-Transducer~\cite{Ryan19}, Transformer-Transducer~\cite{zhang2020transformer}, ConvNet-Transducer~\cite{han2020contextnet} and Conformer-Transducer~\cite{gulati2020conformer}, and achieve 150-300ms latency reduction over previous art without accuracy degradation on a Voice Search test set. FastEmit also improves streaming ASR accuracy from 4.4%/8.9% to 3.1%/7.5% WER, meanwhile reduces 90th percentile latency from 210 ms to only 30 ms on LibriSpeech. View details
    Preview abstract For various speech-related tasks, confidence scores from a speech recogniser are a useful measure to assess the quality of transcriptions. In traditional hidden Markov model-based automatic speech recognition (ASR) systems, confidence scores can be reliably obtained from word posteriors in decoding lattices. However, for an ASR system with an auto-regressive decoder such as an attention-based sequence-to-sequence model, computing word posteriors is difficult. An obvious alternative is to use the decoder softmax probability as the model confidence. To reach good recognition performance, end-to-end ASR models tend to be very large. However, large models can easily memorise training sequences, which results in overestimated confidence scores. Some regularisation techniques can directly affect softmax probabilities. In this paper, we first examine how some commonly used regularisation methods influence the confidence scores and study the overconfident behaviour of end-to-end models. Then we propose a lightweight and effective approach named confidence estimation module (CEM) on top of an existing end-to-end ASR model. Experiments on LibriSpeech show that CEM can mitigate the overconfidence problem and can produce more reliable confidence scores with and without shallow fusion of a language model. Further analysis shows that CEM generalises well to speech from a moderately mismatched domain and can potentially improve downstream tasks such as semi-supervised learning. View details
    Preview abstract In this paper, we propose a solution to allow speaker conditioned speech models, such as VoiceFilter-Lite, to support an arbitrary number of enrolled users in a single pass. This is achieved by using an attention mechanism on multiple speaker embeddings to compute a single attentive embedding, which is then used as a side input to the model. We implemented multi-user VoiceFilter-Lite and evaluated it for two tasks: (1) a standard text-independent speaker verification task, where the input audio may contain overlapped speech; (2) a personalized keyphrase detection task, where ASR has to detect keyphrases from multiple enrolled users in a noisy environment. Our experiments show that with up to four enrolled users, multi-user VoiceFilter-Lite is able to significantly reduce speaker verification errors when there is overlapped speech, without hurting the performance under other acoustic conditions. This attentive speaker embedding approach can also be easily applied to other speaker-conditioned models such as personal VAD and personalized ASR. View details
    Preview abstract End-to-end models that condition the output sequence on all previously predicted labels have emerged as popular alternatives to conventional systems for automatic speech recognition (ASR). Since distinct label histories correspond to distinct models states, such models are decoded using an approximate beam-search 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 decoding efficiency 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, or up to 15.7% when lattice rescoring is applied. View details
    Preview abstract On-device end-to-end (E2E) models have shown improvementsover a conventional model on Search test sets in both quality, as measured by Word Error Rate (WER), and latency, measured by the time the result is finalized after the user stops speaking. However, the E2E model is trained on a small fraction of audio-text pairs compared to the 100 billion text utterances that a conventional language model (LM) is trained with. Thus E2E models perform poorly on rare words and phrases. In this paper, building upon the two-pass streaming Cascaded Encoder E2E model, we explore using a Hybrid Autoregressive Transducer (HAT) factorization to better integrate an on-device neural LM trained on text-only data. Furthermore, to further improve decoder latency we introduce a non-recurrent embedding decoder, in place of the typical LSTM decoder, into the Cascaded Encoder model. Overall, we present a streaming on-device model that incorporates an external neural LM and outperforms the conventional model in both search and rare-word quality, as well as latency, and is 318X smaller. View details
    Preview abstract In this paper, we introduce a streaming keyphrase detection system that can be easily customized to accurately detect any phrase composed of words from a large vocabulary. The system is implemented with an end-to-end trained automatic speech recognition (ASR) model and a text-independent speaker verification model. To address the challenge of detecting these keyphrases under various noisy conditions, a speaker separation model is added to the feature frontend of the speaker verification model, and an adaptive noise cancellation (ANC) algorithm is included to exploit the cross-microphone noise coherence. Our experiments show that the text-independent speaker recognition model largely reduces the false triggering rate of the keyphrase detection, while the speaker separation model and adaptive noise cancellation largely reduce false rejections. View details
    Preview abstract Confidence scores are very useful for downstream applicationsof automatic speech recognition (ASR) systems. Recent workshave proposed using neural attention models to learn word or ut-terance confidence scores for end-to-end (E2E) ASR. By them-selves, word confidence does not model deletions, and utteranceconfidence discards much of the useful word-level training sig-nals. This paper studies the effect of adding utterance-level lossand individual deletion loss to the framework proposed in [1].Empirical results show that multi-task learning with all threeobjectives improves confidence metrics (NCE, AUC, RMSE)without the need for increasing the model size of the trans-former feature extractor. Using the utterance-level confidencefor rescoring also decreases the word error rates on Google’sVoice Search and long-tail datasets by 3-5% relative. View details
    Preview abstract 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. View details
    Preview abstract Two-pass models have achieved better quality for on-device speech recognition, where a 1st-pass recurrent neural network transducer (RNN-T) model generates hypotheses in a streaming fashion, and a 2nd-pass Listen, Attend and Spell (LAS) model re-scores the hypotheses with full audio sequence context. Such models provide both fast responsiveness with the 1st-pass model and better quality with the 2nd-pass model. The computation latency from the 2nd-pass model is a critical problem, as the model has to wait for the speech and hypotheses from the first pass to be complete. Yet the rescoring latency is constrained by the recurrent nature of LSTM, as the processing for each sequence has to run sequentially. In this work we explore replacing the LSTM layers in the 2nd-pass rescorer with Transformer layers, which can process the entire hypothesis sequences in parallel and can therefore utilize the on-device computation resources more efficiently. Compared with an LAS-based baseline, our proposed transformer rescorer achieves more than 50% latency reduction with quality improvement. View details
    Preview abstract End-to-end (E2E) models fold the acoustic, pronunciation and language models of a conventional speech recognition model into one neural network with a much smaller number of parameters than a conventional ASR system, thus making it suitable for on-device applications. For example, Recurrent neural network transducer (RNN-T) as a streaming E2E model that has shown promising potential for on-device ASR. For such applications, quality and latency are two critical factors. We propose to reduce E2E model's latency by extending the RNN-T endpointer (RNN-T EP) model with additional early and late penalties. By further applying the minimum word error rate (MWER) training technique, we achieved 8.0% relative word error rate (WER) reduction and 130ms 90-percentile latency reduction on a Voice search test set. We also experimented with a second pass Listen, Attend and Spell (LAS) rescorer for the RNN-T EP model. Although it cannot directly improve the first pass latency, the large WER reduction actually give us more room to trade WER for latency. RNN-T+LAS, together with EMBR training brings in 17.3% relative WER reduction while maintaining similar 120ms 90-percentile latency reductions. View details
    Preview abstract We introduce VoiceFilter-Lite, a single-channel source separation model that runs on the device to preserve only the speech signals from a target user, as part of a streaming speech recognition system. Delivering such a model presents numerous challenges: It should improve the performance when the input signal consists of overlapped speech, and must not hurt the speech recognition performance under all other acoustic conditions. Besides, this model must be tiny, fast, and perform inference in a streaming fashion, in order to have minimal impact on CPU, memory, battery and latency. We propose novel techniques to meet these multi-faceted requirements, including using a new asymmetric loss, and adopting adaptive runtime suppression strength. We also show that such a model can be quantized as a 8-bit integer model and run in realtime. View details
    Preview abstract Latency is a crucial metric for streaming speech recognition systems. In this paper, we reduce latency by fetching responses early based on the partial recognition results and refer to it as prefetching. Specifically, prefetching works by submitting partial recognition results for subsequent processing such as obtaining assistant server responses or second-pass rescoring before the recognition result is finalized. If the partial result matches the final recognition result, the early fetched response can be delivered to the user instantly. This effectively speeds up the system by saving the execution latency that typically happens after recognition is completed. Prefetching can be triggered multiple times for a single query, but this leads to multiple rounds of downstream processing and increases the computation costs. It is hence desirable to fetch the result sooner but meanwhile limiting the number of prefetches. To achieve the best trade-off between latency and computation cost, we investigated a series of prefetching decision models including decoder silence based prefetching, acoustic silence based prefetching and end-to-end prefetching. In this paper, we demonstrate the proposed prefetching mechanism reduced 200 ms for a system that consists of a streaming first pass model using recurrent neural network transducer (RNN-T) and a non-streaming second pass rescoring model using Listen, Attend and Spell (LAS) [1]. We observe that the endto-end prefetching provides the best trade-off between cost and latency that is 100 ms faster compared to silence based prefetching at a fixed prefetch rate. View details
    Preview abstract Thus far, end-to-end (E2E) models have not 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. View details
    Preview abstract Recently, we introduced a 2-pass on-device E2E model, which runs RNN-T in the first-pass and then rescores/redecodes this with a LAS decoder. This on-device model was similar in performance compared to a state-of-the-art conventional model. However, like many E2E models it is trained on supervised audio-text pairs and thus did poorly on rare-words compared to a conventional model trained on a much larger text-corpora. In this work, we introduce a joint acoustic and text-only decoder (JATD) into the LAS decoder, which allows the LAS decoder to be trained on a much larger text-corporate. We find that the JATD model provides between a 3-10\% relative improvement in WER compared to a LAS decoder trained on only supervised audio-text pairs across a variety of proper noun test sets. View details
    Preview abstract 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. View details
    Preview abstract The tradeoff between word error rate (WER) and latency is very important for online automatic speech recognition (ASR) applications. We want the system to endpoint and close the microphone as quickly as possible, without degrading WER. For conventional ASR systems, endpointing is a separate model from the acoustic, pronunciation and language models (AM, PM, LM), which can often cause endpointer problems, with either a higher WER or larger latency. In going with the all-neural spirit of end-to-end (E2E) models, which fold the AM, PM and LM into one neural network, in this work we look at foldinging the endpointer into the model. On a large vocabulary Voice Search task, we show that joint optimization of the endpointer with the E2E model results in no quality degradation and reduces latency by more than a factor of 2 compared to having a separate endpointer with the E2E model. View details
    Preview abstract 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. However, this model still lags behind a large state-of-the-art conventional model in quality. On the other hand, a non-streaming E2E Listen, Attend and Spell (LAS) model has shown comparable quality to large conventional models. 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. View details
    Preview abstract We develop streaming keyword spotting systems using a recurrent neural network transducer (RNN-T) model: an all-neural, end-to-end trained, sequence-to-sequence model which jointly learns acoustic and language model components. Our models are trained to predict either phonemes or graphemes as subword units, thus allowing us to detect arbitrary keyword phrases, without any out-of-vocabulary words. In order to adapt the models to the requirements of keyword spotting, we propose a novel technique which biases the RNN-T system towards a specific keyword of interest. Our systems are compared against a strong sequence-trained, connectionist temporal classification (CTC) based “keyword-filler” baseline, which is augmented with a separate phoneme language model. Overall, our RNN-T system with the proposed biasing technique significantly improves performance over the baseline system. View details
    No Results Found