Bhuvana Ramabhadran

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.

Gary Wang, Ekin D. Cubuk, Andrew Rosenberg et al. · deepmind

Data augmentation is a ubiquitous technique used to provide robustness to automatic speech recognition (ASR) training. However, even as so much of the ASR training process has become automated and more "end-to-end", the data augmentation policy (what augmentation functions to use, and how to apply them) remains hand-crafted. We present Graph-Augment, a technique to define the augmentation space as directed acyclic graphs (DAGs) and search over this space to optimize the augmentation policy itself. We show that given the same computational budget, policies produced by G-Augment are able to perform better than SpecAugment policies obtained by random search on fine-tuning tasks on CHiME-6 and AMI. G-Augment is also able to establish a new state-of-the-art ASR performance on the CHiME-6 evaluation set (30.7% WER). We further demonstrate that G-Augment policies show better transfer properties across warm-start to cold-start training and model size compared to random-searched SpecAugment policies.

Shikhar Vashishth, Harman Singh, Shikhar Bharadwaj et al. · cmu, deepmind

Representing speech as discrete tokens provides a framework for transforming speech into a format that closely resembles text, thus enabling the use of speech as an input to the widely successful large language models (LLMs). Currently, while several speech tokenizers have been proposed, there is ambiguity regarding the properties that are desired from a tokenizer for specific downstream tasks and its overall generalizability. Evaluating the performance of tokenizers across different downstream tasks is a computationally intensive effort that poses challenges for scalability. To circumvent this requirement, we present STAB (Speech Tokenizer Assessment Benchmark), a systematic evaluation framework designed to assess speech tokenizers comprehensively and shed light on their inherent characteristics. This framework provides a deeper understanding of the underlying mechanisms of speech tokenization, thereby offering a valuable resource for expediting the advancement of future tokenizer models and enabling comparative analysis using a standardized benchmark. We evaluate the STAB metrics and correlate this with downstream task performance across a range of speech tasks and tokenizer choices.

Yochai Blau, Rohan Agrawal, Lior Madmony et al. · deepmind

Accurate recognition of specific categories, such as persons' names, dates or other identifiers is critical in many Automatic Speech Recognition (ASR) applications. As these categories represent personal information, ethical use of this data including collection, transcription, training and evaluation demands special care. One way of ensuring the security and privacy of individuals is to redact or eliminate Personally Identifiable Information (PII) from collection altogether. However, this results in ASR models that tend to have lower recognition accuracy of these categories. We use text-injection to improve the recognition of PII categories by including fake textual substitutes of PII categories in the training data using a text injection method. We demonstrate substantial improvement to Recall of Names and Dates in medical notes while improving overall WER. For alphanumeric digit sequences we show improvements to Character Error Rate and Sentence Accuracy.

Zhehuai Chen, Yu Zhang, Andrew Rosenberg et al.

We present Maestro, a self-supervised training method to unify representations learnt from speech and text modalities. Self-supervised learning from speech signals aims to learn the latent structure inherent in the signal, while self-supervised learning from text attempts to capture lexical information. Learning aligned representations from unpaired speech and text sequences is a challenging task. Previous work either implicitly enforced the representations learnt from these two modalities to be aligned in the latent space through multitasking and parameter sharing or explicitly through conversion of modalities via speech synthesis. While the former suffers from interference between the two modalities, the latter introduces additional complexity. In this paper, we propose Maestro, a novel algorithm to learn unified representations from both these modalities simultaneously that can transfer to diverse downstream tasks such as Automated Speech Recognition (ASR) and Speech Translation (ST). Maestro learns unified representations through sequence alignment, duration prediction and matching embeddings in the learned space through an aligned masked-language model loss. We establish a new state-of-the-art (SOTA) on VoxPopuli multilingual ASR with a 8% relative reduction in Word Error Rate (WER), multidomain SpeechStew ASR (3.7% relative) and 21 languages to English multilingual ST on CoVoST 2 with an improvement of 2.8 BLEU averaged over 21 languages.

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.

Tongzhou Chen, Cyril Allauzen, Yinghui Huang et al.

In this work, we study the impact of Large-scale Language Models (LLM) on Automated Speech Recognition (ASR) of YouTube videos, which we use as a source for long-form ASR. We demonstrate up to 8\% relative reduction in Word Error Eate (WER) on US English (en-us) and code-switched Indian English (en-in) long-form ASR test sets and a reduction of up to 30\% relative on Salient Term Error Rate (STER) over a strong first-pass baseline that uses a maximum-entropy based language model. Improved lattice processing that results in a lattice with a proper (non-tree) digraph topology and carrying context from the 1-best hypothesis of the previous segment(s) results in significant wins in rescoring with LLMs. We also find that the gains in performance from the combination of LLMs trained on vast quantities of available data (such as C4) and conventional neural LMs is additive and significantly outperforms a strong first-pass baseline with a maximum entropy LM. Copyright 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

Takaaki Saeki, Heiga Zen, Zhehuai Chen et al.

This paper proposes Virtuoso, a massively multilingual speech-text joint semi-supervised learning framework for text-to-speech synthesis (TTS) models. Existing multilingual TTS typically supports tens of languages, which are a small fraction of the thousands of languages in the world. One difficulty to scale multilingual TTS to hundreds of languages is collecting high-quality speech-text paired data in low-resource languages. This study extends Maestro, a speech-text joint pretraining framework for automatic speech recognition (ASR), to speech generation tasks. To train a TTS model from various types of speech and text data, different training schemes are designed to handle supervised (paired TTS and ASR data) and unsupervised (untranscribed speech and unspoken text) datasets. Experimental evaluation shows that 1) multilingual TTS models trained on Virtuoso can achieve significantly better naturalness and intelligibility than baseline ones in seen languages, and 2) they can synthesize reasonably intelligible and naturally sounding speech for unseen languages where no high-quality paired TTS data is available.

Zhehuai Chen, Ankur Bapna, Andrew Rosenberg et al.

Training state-of-the-art Automated Speech Recognition (ASR) models typically requires a substantial amount of transcribed speech. In this work, we demonstrate that a modality-matched joint speech and text model can be leveraged to train a massively multilingual ASR model without any supervised (manually transcribed) speech for some languages. This paper explores the use of jointly learnt speech and text representations in a massively multilingual, zero supervised speech, real-world setting to expand the set of languages covered by ASR with only unlabeled speech and text in the target languages. Using the FLEURS dataset, we define the task to cover $102$ languages, where transcribed speech is available in $52$ of these languages and can be used to improve end-to-end ASR quality on the remaining $50$. First, we show that by combining speech representations with byte-level text representations and use of language embeddings, we can dramatically reduce the Character Error Rate (CER) on languages with no supervised speech from 64.8\% to 30.8\%, a relative reduction of 53\%. Second, using a subset of South Asian languages we show that Maestro-U can promote knowledge transfer from languages with supervised speech even when there is limited to no graphemic overlap. Overall, Maestro-U closes the gap to oracle performance by 68.5\% relative and reduces the CER of 19 languages below 15\%.

Gary Wang, Kyle Kastner, Ankur Bapna et al.

Recently, a number of approaches to train speech models by incorpo-rating text into end-to-end models have been developed, with Mae-stro advancing state-of-the-art automatic speech recognition (ASR)and Speech Translation (ST) performance. In this paper, we expandour understanding of the resulting shared speech-text representationswith two types of analyses. First we examine the limits of speech-free domain adaptation, finding that a corpus-specific duration modelfor speech-text alignment is the most important component for learn-ing a shared speech-text representation. Second, we inspect the sim-ilarities between activations of unimodal (speech or text) encodersas compared to the activations of a shared encoder. We find that theshared encoder learns a more compact and overlapping speech-textrepresentation than the uni-modal encoders. We hypothesize that thispartially explains the effectiveness of the Maestro shared speech-textrepresentations.

Kartik Audhkhasi, Yinghui Huang, Bhuvana Ramabhadran et al.

Attention layers are an integral part of modern end-to-end automatic speech recognition systems, for instance as part of the Transformer or Conformer architecture. Attention is typically multi-headed, where each head has an independent set of learned parameters and operates on the same input feature sequence. The output of multi-headed attention is a fusion of the outputs from the individual heads. We empirically analyze the diversity between representations produced by the different attention heads and demonstrate that the heads become highly correlated during the course of training. We investigate a few approaches to increasing attention head diversity, including using different attention mechanisms for each head and auxiliary training loss functions to promote head diversity. We show that introducing diversity-promoting auxiliary loss functions during training is a more effective approach, and obtain WER improvements of up to 6% relative on the Librispeech corpus. Finally, we draw a connection between the diversity of attention heads and the similarity of the gradients of head parameters.

Mohammad Zeineldeen, Kartik Audhkhasi, Murali Karthick Baskar et al.

This work studies knowledge distillation (KD) and addresses its constraints for recurrent neural network transducer (RNN-T) models. In hard distillation, a teacher model transcribes large amounts of unlabelled speech to train a student model. Soft distillation is another popular KD method that distills the output logits of the teacher model. Due to the nature of RNN-T alignments, applying soft distillation between RNN-T architectures having different posterior distributions is challenging. In addition, bad teachers having high word-error-rate (WER) reduce the efficacy of KD. We investigate how to effectively distill knowledge from variable quality ASR teachers, which has not been studied before to the best of our knowledge. We show that a sequence-level KD, full-sum distillation, outperforms other distillation methods for RNN-T models, especially for bad teachers. We also propose a variant of full-sum distillation that distills the sequence discriminative knowledge of the teacher leading to further improvement in WER. We conduct experiments on public datasets namely SpeechStew and LibriSpeech, and on in-house production data.

Gary Wang, Andrew Rosenberg, Bhuvana Ramabhadran et al.

Automatic speech recognition (ASR) needs to be robust to speaker differences. Voice Conversion (VC) modifies speaker characteristics of input speech. This is an attractive feature for ASR data augmentation. In this paper, we demonstrate that voice conversion can be used as a data augmentation technique to improve ASR performance, even on LibriSpeech, which contains 2,456 speakers. For ASR augmentation, it is necessary that the VC model be robust to a wide range of input speech. This motivates the use of a non-autoregressive, non-parallel VC model, and the use of a pretrained ASR encoder within the VC model. This work suggests that despite including many speakers, speaker diversity may remain a limitation to ASR quality. Finally, interrogation of our VC performance has provided useful metrics for objective evaluation of VC quality.

Bolaji Yusuf, Murali Karthick Baskar, Andrew Rosenberg et al.

This paper explores speculative speech recognition (SSR), where we empower conventional automatic speech recognition (ASR) with speculation capabilities, allowing the recognizer to run ahead of audio. We introduce a metric for measuring SSR performance and we propose a model which does SSR by combining a RNN-Transducer-based ASR system with an audio-prefixed language model (LM). The ASR system transcribes ongoing audio and feeds the resulting transcripts, along with an audio-dependent prefix, to the LM, which speculates likely completions for the transcriptions. We experiment with a variety of ASR datasets on which show the efficacy our method and the feasibility of SSR as a method of reducing ASR latency.

Murali Karthick Baskar, Andrew Rosenberg, Bhuvana Ramabhadran et al.

We introduce O-1, a new self-training objective to reduce training bias and unify training and evaluation metrics for speech recognition. O-1 is a faster variant of Expected Minimum Bayes Risk (EMBR), that boosts the oracle hypothesis and can accommodate both supervised and unsupervised data. We demonstrate the effectiveness of our approach in terms of recognition on publicly available SpeechStew datasets and a large-scale, in-house data set. On Speechstew, the O-1 objective closes the gap between the actual and oracle performance by 80\% relative compared to EMBR which bridges the gap by 43\% relative. O-1 achieves 13\% to 25\% relative improvement over EMBR on the various datasets that SpeechStew comprises of, and a 12\% relative gap reduction with respect to the oracle WER over EMBR training on the in-house dataset. Overall, O-1 results in a 9\% relative improvement in WER over EMBR, thereby speaking to the scalability of the proposed objective for large-scale datasets.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Pooneh Mousavi, Gallil Maimon, Adel Moumen et al.

Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks. They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area. For more information, including our main results and tokenizer database, please refer to our website: https://poonehmousavi.github.io/dates-website/.

Christopher Richardson, Roshan Sharma, Neeraj Gaur et al.

Zero-shot domain adaptation for dialogue state tracking (DST) remains a challenging problem in task-oriented dialogue (TOD) systems, where models must generalize to target domains unseen at training time. Current large language model approaches for zero-shot domain adaptation rely on prompting to introduce knowledge pertaining to the target domains. However, their efficacy strongly depends on prompt engineering, as well as the zero-shot ability of the underlying language model. In this work, we devise a novel data augmentation approach, Schema Augmentation, that improves the zero-shot domain adaptation of language models through fine-tuning. Schema Augmentation is a simple but effective technique that enhances generalization by introducing variations of slot names within the schema provided in the prompt. Experiments on MultiWOZ and SpokenWOZ showed that the proposed approach resulted in a substantial improvement over the baseline, in some experiments achieving over a twofold accuracy gain over unseen domains while maintaining equal or superior performance over all domains.

Murali Karthick Baskar, Andrew Rosenberg, Bhuvana Ramabhadran et al.

In this paper, we focus on addressing the constraints faced when applying LLMs to ASR. Recent works utilize prefixLM-type models, which directly apply speech as a prefix to LLMs for ASR. We have found that optimizing speech prefixes leads to better ASR performance and propose applying RNNT loss to perform speech prefix-tuning. This is a simple approach and does not increase the model complexity or alter the inference pipeline. We also propose language-based soft prompting to further improve with frozen LLMs. Empirical analysis on realtime testset from 10 Indic languages demonstrate that our proposed speech prefix-tuning yields improvements with both frozen and fine-tuned LLMs. Our recognition results on an average of 10 Indics show that the proposed prefix-tuning with RNNT loss results in a 12\% relative improvement in WER over the baseline with a fine-tuned LLM. Our proposed approches with the frozen LLM leads to a 31\% relative improvement over basic soft-prompting prefixLM.

Neeraj Gaur, Rohan Agrawal, Gary Wang et al.

This paper introduces ASTRA, a novel method for improving Automatic Speech Recognition (ASR) through text injection.Unlike prevailing techniques, ASTRA eliminates the need for sampling to match sequence lengths between speech and text modalities. Instead, it leverages the inherent alignments learned within CTC/RNNT models. This approach offers the following two advantages, namely, avoiding potential misalignment between speech and text features that could arise from upsampling and eliminating the need for models to accurately predict duration of sub-word tokens. This novel formulation of modality (length) matching as a weighted RNNT objective matches the performance of the state-of-the-art duration-based methods on the FLEURS benchmark, while opening up other avenues of research in speech processing.

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%.

Murali Karthick Baskar, Andrew Rosenberg, Bhuvana Ramabhadran et al.

Masked speech modeling (MSM) methods such as wav2vec2 or w2v-BERT learn representations over speech frames which are randomly masked within an utterance. While these methods improve performance of Automatic Speech Recognition (ASR) systems, they have one major limitation. They treat all unsupervised speech samples with equal weight, which hinders learning as not all samples have relevant information to learn meaningful representations. In this work, we address this limitation. We propose ask2mask (ATM), a novel approach to focus on specific samples during MSM pre-training. ATM employs an external ASR model or \textit{scorer} to weight unsupervised input samples in two different ways: 1) A fine-grained data selection is performed by masking over the highly confident input frames as chosen by the scorer. This allows the model to learn meaningful representations. 2) ATM is further extended to focus at utterance-level by weighting the final MSM loss with the utterance-level confidence score. We conduct fine-tuning experiments on two well-benchmarked corpora: LibriSpeech (matching the pre-training data) and Commonvoice, TED-LIUM, AMI and CHiME-6 (not matching the pre-training data). The results substantiate the efficacy of ATM on significantly improving the recognition performance under mismatched conditions (up to 11.6\% relative over published results and upto 4.46\% relative over our internal baseline) while still yielding modest improvements under matched conditions.

Yu Zhang, Daniel S. Park, Wei Han et al.

We summarize the results of a host of efforts using giant automatic speech recognition (ASR) models pre-trained using large, diverse unlabeled datasets containing approximately a million hours of audio. We find that the combination of pre-training, self-training and scaling up model size greatly increases data efficiency, even for extremely large tasks with tens of thousands of hours of labeled data. In particular, on an ASR task with 34k hours of labeled data, by fine-tuning an 8 billion parameter pre-trained Conformer model we can match state-of-the-art (SoTA) performance with only 3% of the training data and significantly improve SoTA with the full training set. We also report on the universal benefits gained from using big pre-trained and self-trained models for a large set of downstream tasks that cover a wide range of speech domains and span multiple orders of magnitudes of dataset sizes, including obtaining SoTA performance on many public benchmarks. In addition, we utilize the learned representation of pre-trained networks to achieve SoTA results on non-ASR tasks.

Zhehuai Chen, Yu Zhang, Andrew Rosenberg et al.

Self-supervised pretraining for Automated Speech Recognition (ASR) has shown varied degrees of success. In this paper, we propose to jointly learn representations during pretraining from two different modalities: speech and text. The proposed method, tts4pretrain complements the power of contrastive learning in self-supervision with linguistic/lexical representations derived from synthesized speech, effectively learning from untranscribed speech and unspoken text. Lexical learning in the speech encoder is enforced through an additional sequence loss term that is coupled with contrastive loss during pretraining. We demonstrate that this novel pretraining method yields Word Error Rate (WER) reductions of 10% relative on the well-benchmarked, Librispeech task over a state-of-the-art baseline pretrained with wav2vec2.0 only. The proposed method also serves as an effective strategy to compensate for the lack of transcribed speech, effectively matching the performance of 5000 hours of transcribed speech with just 100 hours of transcribed speech on the AMI meeting transcription task. Finally, we demonstrate WER reductions of up to 15% on an in-house Voice Search task over traditional pretraining. Incorporating text into encoder pretraining is complimentary to rescoring with a larger or in-domain language model, resulting in additional 6% relative reduction in WER.

Arindrima Datta, Guanlong Zhao, Bhuvana Ramabhadran et al.

Automated speech recognition coverage of the world's languages continues to expand. However, standard phoneme based systems require handcrafted lexicons that are difficult and expensive to obtain. To address this problem, we propose a training methodology for a grapheme-based speech recognizer that can be trained in a purely data-driven fashion. Built with LSTM networks and trained with the cross-entropy loss, the grapheme-output acoustic models we study are also extremely practical for real-world applications as they can be decoded with conventional ASR stack components such as language models and FST decoders, and produce good quality audio-to-grapheme alignments that are useful in many speech applications. We show that the grapheme models are competitive in WER with their phoneme-output counterparts when trained on large datasets, with the advantage that grapheme models do not require explicit linguistic knowledge as an input. We further compare the alignments generated by the phoneme and grapheme models to demonstrate the quality of the pronunciations learnt by them using four Indian languages that vary linguistically in spoken and written forms.

Arindrima Datta, Bhuvana Ramabhadran, Jesse Emond et al.

Multilingual Automated Speech Recognition (ASR) systems allow for the joint training of data-rich and data-scarce languages in a single model. This enables data and parameter sharing across languages, which is especially beneficial for the data-scarce languages. However, most state-of-the-art multilingual models require the encoding of language information and therefore are not as flexible or scalable when expanding to newer languages. Language-independent multilingual models help to address this issue, and are also better suited for multicultural societies where several languages are frequently used together (but often rendered with different writing systems). In this paper, we propose a new approach to building a language-agnostic multilingual ASR system which transforms all languages to one writing system through a many-to-one transliteration transducer. Thus, similar sounding acoustics are mapped to a single, canonical target sequence of graphemes, effectively separating the modeling and rendering problems. We show with four Indic languages, namely, Hindi, Bengali, Tamil and Kannada, that the language-agnostic multilingual model achieves up to 10% relative reduction in Word Error Rate (WER) over a language-dependent multilingual model.

Guangzhi Sun, Yu Zhang, Ron J. Weiss et al.

Recent neural text-to-speech (TTS) models with fine-grained latent features enable precise control of the prosody of synthesized speech. Such models typically incorporate a fine-grained variational autoencoder (VAE) structure, extracting latent features at each input token (e.g., phonemes). However, generating samples with the standard VAE prior often results in unnatural and discontinuous speech, with dramatic prosodic variation between tokens. This paper proposes a sequential prior in a discrete latent space which can generate more naturally sounding samples. This is accomplished by discretizing the latent features using vector quantization (VQ), and separately training an autoregressive (AR) prior model over the result. We evaluate the approach using listening tests, objective metrics of automatic speech recognition (ASR) performance, and measurements of prosody attributes. Experimental results show that the proposed model significantly improves the naturalness in random sample generation. Furthermore, initial experiments demonstrate that randomly sampling from the proposed model can be used as data augmentation to improve the ASR performance.

Andrew Rosenberg, Yu Zhang, Bhuvana Ramabhadran et al.

Recent success of the Tacotron speech synthesis architecture and its variants in producing natural sounding multi-speaker synthesized speech has raised the exciting possibility of replacing expensive, manually transcribed, domain-specific, human speech that is used to train speech recognizers. The multi-speaker speech synthesis architecture can learn latent embedding spaces of prosody, speaker and style variations derived from input acoustic representations thereby allowing for manipulation of the synthesized speech. In this paper, we evaluate the feasibility of enhancing speech recognition performance using speech synthesis using two corpora from different domains. We explore algorithms to provide the necessary acoustic and lexical diversity needed for robust speech recognition. Finally, we demonstrate the feasibility of this approach as a data augmentation strategy for domain-transfer. We find that improvements to speech recognition performance is achievable by augmenting training data with synthesized material. However, there remains a substantial gap in performance between recognizers trained on human speech those trained on synthesized speech.

Anjuli Kannan, Arindrima Datta, Tara N. Sainath et al.

Multilingual end-to-end (E2E) models have shown great promise in expansion of automatic speech recognition (ASR) coverage of the world's languages. They have shown improvement over monolingual systems, and have simplified training and serving by eliminating language-specific acoustic, pronunciation, and language models. This work presents an E2E multilingual system which is equipped to operate in low-latency interactive applications, as well as handle a key challenge of real world data: the imbalance in training data across languages. Using nine Indic languages, we compare a variety of techniques, and find that a combination of conditioning on a language vector and training language-specific adapter layers produces the best model. The resulting E2E multilingual model achieves a lower word error rate (WER) than both monolingual E2E models (eight of nine languages) and monolingual conventional systems (all nine languages).

Yu Zhang, Ron J. Weiss, Heiga Zen et al.

We present a multispeaker, multilingual text-to-speech (TTS) synthesis model based on Tacotron that is able to produce high quality speech in multiple languages. Moreover, the model is able to transfer voices across languages, e.g. synthesize fluent Spanish speech using an English speaker's voice, without training on any bilingual or parallel examples. Such transfer works across distantly related languages, e.g. English and Mandarin. Critical to achieving this result are: 1. using a phonemic input representation to encourage sharing of model capacity across languages, and 2. incorporating an adversarial loss term to encourage the model to disentangle its representation of speaker identity (which is perfectly correlated with language in the training data) from the speech content. Further scaling up the model by training on multiple speakers of each language, and incorporating an autoencoding input to help stabilize attention during training, results in a model which can be used to consistently synthesize intelligible speech for training speakers in all languages seen during training, and in native or foreign accents.

Xuesong Yang, Kartik Audhkhasi, Andrew Rosenberg et al.

The performance of automatic speech recognition systems degrades with increasing mismatch between the training and testing scenarios. Differences in speaker accents are a significant source of such mismatch. The traditional approach to deal with multiple accents involves pooling data from several accents during training and building a single model in multi-task fashion, where tasks correspond to individual accents. In this paper, we explore an alternate model where we jointly learn an accent classifier and a multi-task acoustic model. Experiments on the American English Wall Street Journal and British English Cambridge corpora demonstrate that our joint model outperforms the strong multi-task acoustic model baseline. We obtain a 5.94% relative improvement in word error rate on British English, and 9.47% relative improvement on American English. This illustrates that jointly modeling with accent information improves acoustic model performance.

Kartik Audhkhasi, Brian Kingsbury, Bhuvana Ramabhadran et al.

Direct acoustics-to-word (A2W) models in the end-to-end paradigm have received increasing attention compared to conventional sub-word based automatic speech recognition models using phones, characters, or context-dependent hidden Markov model states. This is because A2W models recognize words from speech without any decoder, pronunciation lexicon, or externally-trained language model, making training and decoding with such models simple. Prior work has shown that A2W models require orders of magnitude more training data in order to perform comparably to conventional models. Our work also showed this accuracy gap when using the English Switchboard-Fisher data set. This paper describes a recipe to train an A2W model that closes this gap and is at-par with state-of-the-art sub-word based models. We achieve a word error rate of 8.8%/13.9% on the Hub5-2000 Switchboard/CallHome test sets without any decoder or language model. We find that model initialization, training data order, and regularization have the most impact on the A2W model performance. Next, we present a joint word-character A2W model that learns to first spell the word and then recognize it. This model provides a rich output to the user instead of simple word hypotheses, making it especially useful in the case of words unseen or rarely-seen during training.

Gakuto Kurata, Bhuvana Ramabhadran, George Saon et al.

Language models (LMs) based on Long Short Term Memory (LSTM) have shown good gains in many automatic speech recognition tasks. In this paper, we extend an LSTM by adding highway networks inside an LSTM and use the resulting Highway LSTM (HW-LSTM) model for language modeling. The added highway networks increase the depth in the time dimension. Since a typical LSTM has two internal states, a memory cell and a hidden state, we compare various types of HW-LSTM by adding highway networks onto the memory cell and/or the hidden state. Experimental results on English broadcast news and conversational telephone speech recognition show that the proposed HW-LSTM LM improves speech recognition accuracy on top of a strong LSTM LM baseline. We report 5.1% and 9.9% on the Switchboard and CallHome subsets of the Hub5 2000 evaluation, which reaches the best performance numbers reported on these tasks to date.

Kartik Audhkhasi, Bhuvana Ramabhadran, George Saon et al.

Recent work on end-to-end automatic speech recognition (ASR) has shown that the connectionist temporal classification (CTC) loss can be used to convert acoustics to phone or character sequences. Such systems are used with a dictionary and separately-trained Language Model (LM) to produce word sequences. However, they are not truly end-to-end in the sense of mapping acoustics directly to words without an intermediate phone representation. In this paper, we present the first results employing direct acoustics-to-word CTC models on two well-known public benchmark tasks: Switchboard and CallHome. These models do not require an LM or even a decoder at run-time and hence recognize speech with minimal complexity. However, due to the large number of word output units, CTC word models require orders of magnitude more data to train reliably compared to traditional systems. We present some techniques to mitigate this issue. Our CTC word model achieves a word error rate of 13.0%/18.8% on the Hub5-2000 Switchboard/CallHome test sets without any LM or decoder compared with 9.6%/16.0% for phone-based CTC with a 4-gram LM. We also present rescoring results on CTC word model lattices to quantify the performance benefits of a LM, and contrast the performance of word and phone CTC models.

George Saon, Gakuto Kurata, Tom Sercu et al.

One of the most difficult speech recognition tasks is accurate recognition of human to human communication. Advances in deep learning over the last few years have produced major speech recognition improvements on the representative Switchboard conversational corpus. Word error rates that just a few years ago were 14% have dropped to 8.0%, then 6.6% and most recently 5.8%, and are now believed to be within striking range of human performance. This then raises two issues - what IS human performance, and how far down can we still drive speech recognition error rates? A recent paper by Microsoft suggests that we have already achieved human performance. In trying to verify this statement, we performed an independent set of human performance measurements on two conversational tasks and found that human performance may be considerably better than what was earlier reported, giving the community a significantly harder goal to achieve. We also report on our own efforts in this area, presenting a set of acoustic and language modeling techniques that lowered the word error rate of our own English conversational telephone LVCSR system to the level of 5.5%/10.3% on the Switchboard/CallHome subsets of the Hub5 2000 evaluation, which - at least at the writing of this paper - is a new performance milestone (albeit not at what we measure to be human performance!). On the acoustic side, we use a score fusion of three models: one LSTM with multiple feature inputs, a second LSTM trained with speaker-adversarial multi-task learning and a third residual net (ResNet) with 25 convolutional layers and time-dilated convolutions. On the language modeling side, we use word and character LSTMs and convolutional WaveNet-style language models.

Kartik Audhkhasi, Andrew Rosenberg, Abhinav Sethy et al.

End-to-end (E2E) systems have achieved competitive results compared to conventional hybrid hidden Markov model (HMM)-deep neural network based automatic speech recognition (ASR) systems. Such E2E systems are attractive due to the lack of dependence on alignments between input acoustic and output grapheme or HMM state sequence during training. This paper explores the design of an ASR-free end-to-end system for text query-based keyword search (KWS) from speech trained with minimal supervision. Our E2E KWS system consists of three sub-systems. The first sub-system is a recurrent neural network (RNN)-based acoustic auto-encoder trained to reconstruct the audio through a finite-dimensional representation. The second sub-system is a character-level RNN language model using embeddings learned from a convolutional neural network. Since the acoustic and text query embeddings occupy different representation spaces, they are input to a third feed-forward neural network that predicts whether the query occurs in the acoustic utterance or not. This E2E ASR-free KWS system performs respectably despite lacking a conventional ASR system and trains much faster.

Dmitriy Serdyuk, Kartik Audhkhasi, Philémon Brakel et al.

Modern automatic speech recognition (ASR) systems need to be robust under acoustic variability arising from environmental, speaker, channel, and recording conditions. Ensuring such robustness to variability is a challenge in modern day neural network-based ASR systems, especially when all types of variability are not seen during training. We attempt to address this problem by encouraging the neural network acoustic model to learn invariant feature representations. We use ideas from recent research on image generation using Generative Adversarial Networks and domain adaptation ideas extending adversarial gradient-based training. A recent work from Ganin et al. proposes to use adversarial training for image domain adaptation by using an intermediate representation from the main target classification network to deteriorate the domain classifier performance through a separate neural network. Our work focuses on investigating neural architectures which produce representations invariant to noise conditions for ASR. We evaluate the proposed architecture on the Aurora-4 task, a popular benchmark for noise robust ASR. We show that our method generalizes better than the standard multi-condition training especially when only a few noise categories are seen during training.

Ewout van den Berg, Bhuvana Ramabhadran, Michael Picheny

In this work we study variance in the results of neural network training on a wide variety of configurations in automatic speech recognition. Although this variance itself is well known, this is, to the best of our knowledge, the first paper that performs an extensive empirical study on its effects in speech recognition. We view training as sampling from a distribution and show that these distributions can have a substantial variance. These results show the urgent need to rethink the way in which results in the literature are reported and interpreted.

Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran

We propose Diverse Embedding Neural Network (DENN), a novel architecture for language models (LMs). A DENNLM projects the input word history vector onto multiple diverse low-dimensional sub-spaces instead of a single higher-dimensional sub-space as in conventional feed-forward neural network LMs. We encourage these sub-spaces to be diverse during network training through an augmented loss function. Our language modeling experiments on the Penn Treebank data set show the performance benefit of using a DENNLM.

Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran et al.

Diversity or complementarity of experts in ensemble pattern recognition and information processing systems is widely-observed by researchers to be crucial for achieving performance improvement upon fusion. Understanding this link between ensemble diversity and fusion performance is thus an important research question. However, prior works have theoretically characterized ensemble diversity and have linked it with ensemble performance in very restricted settings. We present a generalized ambiguity decomposition (GAD) theorem as a broad framework for answering these questions. The GAD theorem applies to a generic convex ensemble of experts for any arbitrary twice-differentiable loss function. It shows that the ensemble performance approximately decomposes into a difference of the average expert performance and the diversity of the ensemble. It thus provides a theoretical explanation for the empirically-observed benefit of fusing outputs from diverse classifiers and regressors. It also provides a loss function-dependent, ensemble-dependent, and data-dependent definition of diversity. We present extensions of this decomposition to common regression and classification loss functions, and report a simulation-based analysis of the diversity term and the accuracy of the decomposition. We finally present experiments on standard pattern recognition data sets which indicate the accuracy of the decomposition for real-world classification and regression problems.

Tara N. Sainath, Lior Horesh, Brian Kingsbury et al.

Hessian-free training has become a popular parallel second or- der optimization technique for Deep Neural Network training. This study aims at speeding up Hessian-free training, both by means of decreasing the amount of data used for training, as well as through reduction of the number of Krylov subspace solver iterations used for implicit estimation of the Hessian. In this paper, we develop an L-BFGS based preconditioning scheme that avoids the need to access the Hessian explicitly. Since L-BFGS cannot be regarded as a fixed-point iteration, we further propose the employment of flexible Krylov subspace solvers that retain the desired theoretical convergence guarantees of their conventional counterparts. Second, we propose a new sampling algorithm, which geometrically increases the amount of data utilized for gradient and Krylov subspace iteration calculations. On a 50-hr English Broadcast News task, we find that these methodologies provide roughly a 1.5x speed-up, whereas, on a 300-hr Switchboard task, these techniques provide over a 2.3x speedup, with no loss in WER. These results suggest that even further speed-up is expected, as problems scale and complexity grows.

Tara N. Sainath, Brian Kingsbury, Abdel-rahman Mohamed et al.

Deep Convolutional Neural Networks (CNNs) are more powerful than Deep Neural Networks (DNN), as they are able to better reduce spectral variation in the input signal. This has also been confirmed experimentally, with CNNs showing improvements in word error rate (WER) between 4-12% relative compared to DNNs across a variety of LVCSR tasks. In this paper, we describe different methods to further improve CNN performance. First, we conduct a deep analysis comparing limited weight sharing and full weight sharing with state-of-the-art features. Second, we apply various pooling strategies that have shown improvements in computer vision to an LVCSR speech task. Third, we introduce a method to effectively incorporate speaker adaptation, namely fMLLR, into log-mel features. Fourth, we introduce an effective strategy to use dropout during Hessian-free sequence training. We find that with these improvements, particularly with fMLLR and dropout, we are able to achieve an additional 2-3% relative improvement in WER on a 50-hour Broadcast News task over our previous best CNN baseline. On a larger 400-hour BN task, we find an additional 4-5% relative improvement over our previous best CNN baseline.