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 smart-phones, 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 5M 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 trans-formations 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 5M parameter model.

Recently, Feature-wise Linear Modulation (FiLM) has been shown to outperform other approaches to incorporate speaker embedding into speech separation and VoiceFilter models. We propose an improved method of incorporating such embeddings into a Voice- Filter frontend for automatic speech recognition (ASR) and text- independent speaker verification (TI-SV). We extend the widely- used Conformer architecture to construct a FiLM Block with additional feature processing before and after the FiLM layers. Apart from its application to single-user VoiceFilter, we show that our system can be easily extended to multi-user VoiceFilter models via element-wise max pooling of the speaker embeddings in a projected space. The final architecture, which we call Conditional Conformer, tightly integrates the speaker embeddings into a Conformer backbone. We improve TI-SV equal error rates by as much as 56% over prior multi-user VoiceFilter models, and our element-wise max pooling reduces relative WER compared to an attention mechanism by as much as 10%.

We propose a multi-output joint network architecture for RNN-T transducer, for multi-task modeling of ASR and auxiliary tasks that rely on ASR outputs. Each output of the joint network predicts tar-get labels with disjoint vocabularies for each task, while sharing the same audio features by the encoder and language model features by the prediction network. Each task is trained with an RNN-T loss that marginalizes over all possible paths, and we allow multiple tasks to share the blank logit so that they are synchronized. We demonstrate our method on two auxiliary tasks, namely capitalization and pause prediction, and discuss different considerations for modeling and inference procedures. For capitalization, we successfully distill capitalization labels from a standalone text normalization model, and achieve competitive Uppercase Error Rate (UER) while offering streaming capability and improved inference efficiency. In addition, our model has similar capitalization accuracy compared to a mixed-case ASR model, but obtains improved WERs if integrated with external language models. For pause prediction, we achieve the same performance as the previous two-step approach while providing a simpler training recipe without affecting ASR accuracy.

In this paper, we explore reducing computational latency of the 2-pass cascaded encoder model [1]. Specifically, we experiment with reducing the size of the causal 1st-pass and adding capacity to the non-causal 2nd-pass, such that the overall latency can be reduced without loss of quality. In addition, we explore using a confidence model for deciding to stop 2nd-pass recognition if we are confident in the 1st-pass hypothesis. Overall, we are able to reduce latency by a factor of 1.7X, compared to the baseline cascaded encoder from [1]. Secondly, with the added capacity in the non-causal 2nd-pass, we find that we can improve WER by up to 7% relative using wav2vec and minimum word-error-rate (MWER) training.

Personalization of on-device speech recognition (ASR) has seen explosive growth in recent years, largely due to the increasing popularity of personal assistant features on mobile devices and smart home speakers. In this work, we present Personal VAD 2.0, a personalized voice activity detector that detects the voice activity of a target speaker, as part of a streaming on-device ASR system. Although previous proof-of-concept studies have validated the effectiveness of Personal VAD, there are still several critical challenges to address before this model can be used in production: first, the quality must be satisfactory in both enrollment and enrollment-less scenarios; second, it should operate in a streaming fashion; and finally, the model size should be small enough to fit a limited latency and CPU/Memory budget. To meet the multi-faceted requirements, we propose a series of novel designs: 1) advanced speaker embedding modulation methods; 2) a new training paradigm to generalize to enrollment-less conditions; 3) architecture and runtime optimizations for latency and resource restrictions. Extensive experiments on a realistic speech recognition system demonstrated the state-of-the-art performance of our proposed method.

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.

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 via 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 three tasks: (1) a streaming automatic speech recognition (ASR) task; (2) a text-independent speaker verification task; and (3) 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 speech recognition and speaker verification errors when there is overlapping speech, without affecting performance under other acoustic conditions. This attentive speaker embedding approach can also be easily applied to other speaker-conditioned models such as personal voice activity detection (VAD) and personalized ASR.

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.

Confidence scores are very useful for downstream applications of automatic speech recognition (ASR) systems. Recent works have proposed using neural networks to learn word or utterance confidence scores for end-to-end ASR. In those studies, word confidence by itself does not model deletions, and utterance confidence does not take advantage of word-level training signals. This paper proposes to jointly learn word confidence, word deletion, and utterance confidence. Empirical results show that multi-task learning with all three objectives improves confidence metrics (NCE, AUC, RMSE) without the need for increasing the model size of the confidence estimation module. Using the utterance-level confidence for rescoring also decreases the word error rates on Google's Voice Search and Long-tail Maps datasets by 3-5% relative, without needing a dedicated neural rescorer.

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 cross-microphone noise coherence. Our experiments show that the text-independent speaker verification model largely reduces the false triggering rate of the keyphrase detection, while the speaker separation model and adaptive noise cancellation largely reduce false rejections.

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 [1] 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.

The demand for fast and accurate incremental speech recognition increases as the applications of automatic speech recognition (ASR) proliferate.Incremental speech recognizers output chunks of partially recognized words while the user is still talking.Partial results can be revised before the ASR finalizes its hypothesis, causing instability issues.We analyze the quality and stability of on-device streaming end-to-end (E2E) ASR models.We first introduce a novel set of metrics that quantify the instability at word and segment levels.We study the impact of several model training techniques that improve E2E model qualities but degrade model stability.We categorize the causes of instability and explore various solutions to mitigate them in a streaming E2E ASR system.

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 recog-nizer 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.

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.

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.

We describe a large vocabulary speech recognition system that is accurate, has low latency, and yet has a small enough memory and computational footprint to run faster than real-time on a Nexus 5 Android smartphone. We employ a quantized Long Short-Term Memory (LSTM) acoustic model trained with connectionist temporal classification (CTC) to directly predict phoneme targets, and further reduce its memory footprint using an SVD-based compression scheme. Additionally, we minimize our memory footprint by using a single language model for both dictation and voice command domains, constructed using Bayesian interpolation. Finally, in order to properly handle device-specific information, such as proper names and other context-dependent information, we inject vocabulary items into the decoder graph and bias the language model on-the-fly. Our system achieves 13.5% word error rate on an open-ended dictation task, running with a median speed that is seven times faster than real-time.

We study the problem of compressing recurrent neural networks (RNNs). In particular, we focus on the compression of RNN acoustic models, which are motivated by the goal of building compact and accurate speech recognition systems which can be run efficiently on mobile devices. In this work, we present a technique for general recurrent model compression that jointly compresses both recurrent and non-recurrent inter-layer weight matrices. We find that the proposed technique allows us to reduce the size of our Long Short-Term Memory (LSTM) acoustic model to a third of its original size with negligible loss in accuracy.

User interactions with mobile devices increasingly depend on voice as a primary input modality. Due to the disadvantages of sending audio across potentially spotty network connections for speech recognition, in recent years there has been growing attention to performing recognition on-device. The limited computational resources, however, typically require additional model constraints. In this work, we explore the task of on-device utterance verification, wherein the recognizer must transcribe an utterance if it is in a target set or reject it as being out of domain. We present a data-driven methodology for mining tens of thousands of target phrases from an existing corpus. We then compare two common garbage-modeling approaches to utterance verification: a sub-word rejection model and a white-listed n-gram model. We examine a deficiency of the sub-word modeling approach and introduce a novel modification that makes use of common prefixes between targeted phrases and non-targeted phrases. We show good performance in the trade-off between recall and word error rate using both the prefix and white-listed n-gram approaches. Finally, we evaluate the prefix-based approach in a hybrid setting where rejected instances are sent to a server-side recognizer.

Speech recognition is an increasingly important input modality, especially for mobile computing. Because errors are unavoidable in real applications, efficient correction methods can greatly enhance the user experience. In this paper we study a reranking and classification strategy for choosing word alternates to display to the user in the framework of a tap-to-correct interface. By employing a logistic regression model to estimate the probability that an alternate will offer a useful correction to the user, we can significantly reduce the average length of the alternates lists generated with no reduction in the number of words they are able to correct.

We summarize the accomplishments of a multi-disciplinary workshop exploring the computational and scientific issues surrounding zero resource (unsupervised) speech technologies and related models of early language acquisition. Centered around the tasks of phonetic and lexical discovery, we consider unified evaluation metrics, present two new approaches for improving speaker independence in the absence of supervision, and evaluate the application of Bayesian word segmentation algorithms to automatic subword unit tokenizations. Finally, we present two strategies for integrating zero resource techniques into supervised settings, demonstrating the potential of unsupervised methods to improve mainstream technologies.

In many ways, the lexicon remains the Achilles heel of modern automatic speech recognizers. Unlike stochastic acoustic and language models that learn the values of their parameters from training data, the baseform pronunciations of words in a recognizer's lexicon are typically specified manually, and do not change, unless they are edited by an expert. Our work presents a novel generative framework that uses speech data to learn stochastic lexicons, thereby taking a step towards alleviating the need for manual intervention and automatically learning high-quality pronunciations for words. We test our model on continuous speech in a weather information domain. In our experiments, we see significant improvements over a manually specified “expert-pronunciation” lexicon. We then analyze variations of the parameter settings used to achieve these gains.

Online streaming companies such as Netflix have become dominant in the media distribution sector. However, such media delivery services often support very rudimentary search, especially for natural language queries. To provide a more natural search interface, we have developed a conversational movie search system, which parses the recognition hypothesis of a spoken query into semantic classes using conditional random fields (CRFs), and then searches an indexed database with the identified semantics. Topic modeling on user-generated content (e.g., movie reviews) is employed for query expansion. Thirteen searching schemas are supported (such as genre, plot, character and soundtrack search). A crowd-sourcing platform was utilized to automatically collect large-scale annotated data for incremental CRF training. Index Terms: conditional random fields, spoken dialogue system 1.

Spoken language systems often rely on static speech recognizers. When the underlying models are improved on-the-fly, training is usually performed using unsupervised methods. In this work, we explore an alternative approach that uses human computation to provide crowd-supervised training of a deployed system. Although the framework we describe is applicable to any stochastic model for which the training data can be generated by non-experts, we demonstrate its utility on the lexicon and language model of a speech recognizer in a cinema voicesearch domain. We show how an initially shaky system can achieve over a 10 % absolute improvement in word error rate (WER) – entirely without expert intervention. We then analyze how these gains were made. 1.

Many speech user interfaces can be improved by incrementally displaying or interpreting a speech recognizer’s current best path as a user speaks. This gives rise to a problem of instability, whereby the best path may change frequently, particularly with respect to the words most recently spoken. Introducing a lag between the audio most recently processed and the portion of the best path shown to the user can lead to more usable incremental results. In the ideal case, the lag introduced would vary to recover exactly the longest stable prefix of the best path. In this paper, we introduce a framework for estimating a stability statistic for each word, and explore the tradeoff of stability and lag by thresholding stability statistics estimated using a variety of features. 1.

This paper explores the use of continuous speech data to learn stochastic lexicons. Building on previous work in which we augmented graphones with acoustic examples of isolated words, we extend our pronunciation mixture model framework to two domains containing spontaneous speech: a weather information retrieval spoken dialogue system and the academic lectures domain. We find that our learned lexicons out-perform expert, hand-crafted lexicons in each domain. Index Terms: grapheme-to-phoneme conversion, pronunciation models, lexical representation

Typically data collection, transcription, language model generation, and deployment are separate phases of creating a spoken language interface. An unfortunate consequence of this is that the recognizer usually remains a static element of systems often deployed in dynamic environments. By providing an API for human intelligence, Amazon Mechanical Turk changes the way system developers can construct spoken language systems. In this work, we describe an architecture that automates and connects these four phases, effectively allowing the developer to grow a spoken language interface. In particular, we show that a human-in-the-loop programming paradigm, in which workers transcribe utterances behind the scenes, can alleviate the need for expert guidance in language model construction. We demonstrate the utility of these organic language models in a voice-search interface for photographs.

A lexicon containing explicit mappings between words and pronunciations is an integral part of most automatic speech recognizers (ASRs). While many ASR components can be trained or adapted using data, the lexicon is one of the few that typically remains static until experts make manual changes. This work takes a step towards alleviating the need for manual intervention by integrating a popular grapheme-to-phoneme conversion technique with acoustic examples to automatically learn highquality baseform pronunciations for unknown words. We explore two models in a Bayesian framework, and discuss their individual advantages and shortcomings. We show that both are able to generate better-than-expert pronunciations with respect to word error rate on an isolated word recognition task. Index Terms: grapheme-to-phoneme conversion, pronunciation models, lexical representation

We explore a new approach to collecting and transcribing speech data by using online educational games. One such game, Voice Race, elicited over 55,000 utterances over a 22 day period, representing 18.7 hours of speech. Voice Race was designed such that the transcripts for a significant subset of utterances can be automatically inferred using the contextual constraints of the game. Game context can also be used to simplify transcription to a multiple choice task, which can be performed by non-experts. We found that one third of the speech collected with Voice Race could be automatically transcribed with over 98 % accuracy; and that an additional 49 % could be labeled cheaply by Amazon Mechanical Turk workers. We demonstrate the utility of the self-labeled speech in an acoustic model adaptation task, which resulted in a reduction in the Voice Race utterance error rate. The collected utterances cover a wide variety of vocabulary, and should be useful across a range of research. Index Terms: speech recognition, data collection, games with a purpose, self-labeled data

Many compelling multimodal prototypes have been developed which pair spoken input and output with a graphical user interface, yet it has often proved difficult to make them available to a large audience. This unfortunate reality limits the degree to which authentic user interactions with such systems can be collected and subsequently analyzed. We present the WAMI toolkit, which alleviates this difficulty by providing a framework for developing, deploying, and evaluating Web-Accessible Multimodal Interfaces in which users interact using speech, mouse, pen, and/or touch. The toolkit makes use of modern web-programming techniques, enabling the development of browser-based applications which rival the quality of traditional native interfaces, yet are available on a wide array of Internet-connected devices. We will showcase several sophisticated multimodal applications developed and deployed using the toolkit, which are available via desktop, laptop, and tablet PCs, as well as via several mobile devices. In addition, we will discuss resources provided by the toolkit for collecting, transcribing, and annotating usage data from multimodal user interactions.