Wen Wu, Bo Li, Chao Zhang, Chung-Cheng Chiu, Qiujia Li, Junwen Bai, Tara Sainath, Phil Woodland. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The end-to-end ASR model is often desired in the streaming multilingual scenario since it is easier to deploy and can benefit from pre-trained speech models such as powerful foundation models. Meanwhile, the heterogeneous nature and imbalanced data abundance of different languages may cause performance degradation, leading to asynchronous peak performance for different languages during training, especially on tail ones. Sometimes even the data itself may become unavailable as a result of the enhanced privacy protection. Existing work tends to significantly increase the model size or learn language-specific decoders to accommodate each language separately. In this study, we explore simple yet effective Language-Dependent Adapter (LDA) finetuning under a cascaded Conformer transducer framework enhanced by teacher pseudolabeling for tail languages in the streaming multilingual ASR. The adapter only accounts for 0.4% of the full model per language. It is plugged into the frozen foundation model and is the only trainable module during the finetuning process with noisy student training. The final model merges the adapter parameters from different checkpoints for different languages. The model performance is validated on a challenging multilingual dictation dataset, which includes 39 tail languages across Latin, Greek, Arabic, etc. Our proposed method brings 12.2% word error rate reduction on average and up to 37.5% on a single locale. Furthermore, we show that our parameter-efficient LDA can match the quality of the full model finetuning, thus greatly alleviating the asynchronous peak performance issue.

End-to-end automatic speech recognition (ASR) models have seen revolutionary quality gains with the recent development of large-scale universal speech models (USM). However, deploying these massive USMs is extremely expensive due to the enormous memory usage and computational cost. Therefore, model compression is an important research topic to fit USM-based ASR under budget in real-world scenarios. In this study, we propose a USM fine-tuning approach for ASR, with a low-bit quantization and N:M structured sparsity aware paradigm on the model weights, reducing the model complexity from parameter precision and matrix topology perspectives. We conducted extensive experiments with a 2-billion parameter USM on a large-scale voice search dataset to evaluate our proposed method. A series of ablation studies validate the effectiveness of up to int4 quantization and 2:4 sparsity. However, a single compression technique fails to recover the performance well under extreme setups including int2 quantization and 1:4 sparsity. By contrast, our proposed method can compress the model to have 9.4% of the size, at the cost of only 7.3% relative word error rate (WER) regressions. We also provided in-depth analyses on the results and discussions on the limitations and potential solutions, which would be valuable for future studies.

Automatic speech recognition (ASR) systems have been shown to have large quality disparities between the language varieties they are intended or expected to recognize. One way to mitigate this is to train or fine-tune models with more representative datasets. But this approach can be hindered by limited in-domain data for training and evaluation. We propose a new way to improve the robustness of a US English short-form speech recognizer using a small amount of out-of-domain (long-form) African American English (AAE) data. We use CORAAL, YouTube and Mozilla Common Voice to train an audio classifier to approximately output whether an utterance is AAE or some other variety including Mainstream American English (MAE). By combining the classifier output with coarse geographic information, we can select a subset of utterances from a large corpus of untranscribed short-form queries for semi-supervised learning at scale. Fine-tuning on this data results in a 38.5% relative word error rate disparity reduction between AAE and MAE without reducing MAE quality.

In the era of large models, the autoregressive nature of decoding often results in latency serving as a significant bottleneck. We propose a non-autoregressive LM-fused ASR system that effectively leverages the parallelization capabilities of accelerator hardware. Our approach combines the Universal Speech Model (USM) and the PaLM 2 language model in per-segment scoring mode, achieving an average relative WER improvement across all languages of 10.8% on FLEURS and 3.6% on YouTube captioning. Furthermore, our comprehensive ablation study analyzes key parameters such as LLM size, context length, vocabulary size, fusion methodology. For instance, we explore the impact of LLM size ranging from 128M to 340B parameters on ASR performance. This study provides valuable insights into the factors influencing the effectiveness of practical large-scale LM-fused speech recognition systems.

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

Online speech recognition, where the model only accesses context to the left, is an important and challenging use case for ASR systems. In this work, we investigate augmenting neural encoders for online ASR by incorporating structured state-space sequence models (S4), a family of models that provide a parameter-efficient way of accessing arbitrarily long left context. We performed systematic ablation studies to compare variants of S4 models and propose two novel approaches that combine them with convolutions. We found that the most effective design is to stack a small S4 using real-valued recurrent weights with a local convolution, allowing them to work complementarily. Our best model achieves WERs of 4.01%/8.53% on test sets from Librispeech, outperforming Conformers with extensively tuned convolution.

A recent paradigm shift in artificial intelligence has seen the rise of foundation models, such as the large language models and the universal speech models. With billions of model parameters and trained with a wide range of data, these foundation models are expected to have a better generalization to different downstream tasks. Efficient adaptation is the key to leveraging these foundation models in a new task or domain. In this paper, we compare several popular parameter-efficient tuning methods, such as vector adaptation, residual adapters, low-rank adapter (LoRA) and prompt-tuning, for automatic speech recognition (ASR) domain adaptation. We use the connectionist temporal classification (CTC) model with Conformer encoder and fused it with a universal language model. We study the effect of adapting either or both of the Conformer encoder and the universal language model. We carry out extensive experiments to study these methods under different hyper-parameter settings and the effect of combining some of these methods. We find that combining vector adaptation and residual adapters with increasing bottleneck dimension achieved the best performance.

Weiran Wang, Rohit Prabhavalkar, Haozhe Shan, Zhong Meng, Dongseong Hwang, Qiujia Li, Khe Chai Sim, Bo Li, James Qin, Xingyu Cai, Adam Stooke, Chengjian Zheng, Yanzhang He, Tara Sainath, Pedro Moreno Mengibar. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

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

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

In this paper, we explore various frame rate reduction schemes on the two-pass cascaded encoder model to improve its efficiency without scarifying the transcription quality. We conduct extensive studies on frame rate reduction strategies, left and right context window length, trade-offs in quality, latency, computation and power consumption, and performance in short-and long-form datasets. With the proposed schemes, we can lower the 2nd pass frame rate to $120 \mathrm{~ms}$, half of the 1st pass’s. This achieves $20 \%$ RTF reduction / $13 \%$ power saving / $19 \%$ lower final latency, without impact on the word-error-rate nor partial results’ latency. If allowing partial latency increase, we can further reduce the frame rate to $180 \mathrm{~ms}$ or even $240 \mathrm{~ms}$ from the 1st pass, and obtain $45 \%$ RTF / 35% power savings, with a similar or even better (on the short-form testset) recognition accuracy.

We investigate using large language models (LLMs) to generate text-only training data for improving multilingual and code-switching automatic speech recognition (ASR) through a text injection method. In a multilingual setup or a low-resource scenario such as code-switching, we propose to generate text data using the state-of-the-art PaLM 2. To better match the generated text data with specific tasks, we use prompt tuning to adapt PaLM 2 to generate domain-relevant multilingual or code-switched text data for text injection. We can achieve significant improvements in Word Error Rate (WER) in both multilingual and code-switching scenarios. The multilingual experiment shows a $6.2 \%$ relative WER reduction on average, i.e., from $11.25 \%$ to $10.55 \%$, compared to a baseline without text injection. The improvement is up to $23.1 \%$ improvement for certain languages. While in the code-switching scenario, we use English-only prompts to generate Mandarin-English code-switching text and achieve a 3.6% relative WER reduction for a code-switching test set, as well as WER reductions in both English and Mandarin monolingual scenarios, $5.3 \%$ and $8.5 \%$ relative, respectively. Our findings demonstrate that leveraging LLMs for text generation and then injection benefits multilingual or code-switching ASR tasks.

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

Typical automatic speech recognition (ASR) systems are built to recognize independent utterances without using the cross-utterance context. However, the context over multiple utterances often provides useful information for the ASR task. In this work, we propose a context-aware end-to-end ASR model that injects the self-attentive context embedding into the decoder of the recurrent neural network transducer (RNN-T). We also propose a factorised 3-way tensor fusion approach to fuse the context embedding with the acoustic representations extracted from the acoustic encoder and the text representations obtained using the prediction network based on the previous subword units. Experimental results on a long-form Youtube ASR task shows that the proposed approach achieves 10.8% relative word error rate reductions.

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.

While large language models (LLM) have made impressive progress in natural language processing, it remains unclear how to utilize them in improving automatic speech recognition (ASR). In this work, we propose to train a single multilingual language model (LM) for shallow fusion in multiple languages. We push the limits of the multilingual LM to cover up to 84 languages by scaling up using a mixture-of-experts LLM, i.e., generalist language model (GLaM). When the number of experts increases, GLaM dynamically selects only two at each decoding step to keep the inference computation roughly constant. We then apply GLaM to a multilingual shallow fusion task based on a state-of-the-art end-to-end model. Compared to a dense LM of similar computation during inference, GLaM reduces the WER of an English long-tail test set by 4.4% relative. In a multilingual shallow fusion task, GLaM improves 41 out of 50 languages with an average relative WER reduction of 3.85%, and a maximum reduction of 10%. Compared to the baseline model, GLaM achieves an average WER reduction of 5.53% over 43 languages.

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

Self-supervised pre-training of a speech foundation model, followed by supervised fine-tuning, has shown impressive quality improvements on automatic speech recognition (ASR) tasks. Fine-tuning separate foundation models for many downstream tasks are expensive since the foundation model is usually very big. Parameter-efficient fine-tuning methods (e.g. adapter, sparse update methods) offer an alternative paradigm where a small set of parameters are updated to adapt the foundation model to new tasks. However, these methods still suffer from a high computational memory cost and slow training speed because they require backpropagation through the entire neural network at each step. In the paper, we analyze the performance of features at different layers of a foundation model on the speech recognition task and propose a novel hierarchical feature fusion method for resource-efficient transfer learning from speech foundation models. Experimental results show that the proposed method can achieve better performance on speech recognition task than existing algorithms with fewer number of trainable parameters, less computational memory cost and faster training speed. After combining with Adapters at all layers, the proposed method can achieve the same performance as fine-tuning the whole model with 97% fewer trainable encoder parameters and 53% faster training speed.

Foundation models (FMs), that are trained on broad data at scale and are adaptable to a wide range of downstream tasks, have brought large interest in the research community. Benefiting from the diverse data sources such as different modalities, languages and application domains, foundation models have demonstrated strong generalization and knowledge transfer capabilities. In this paper, we present a pioneering study towards building an efficient solution for FM-based speech recognition systems. We adopt the recently developed self-supervised BEST-RQ for pretraining, and extend the joint training strategy JUST Hydra for finetuning using both source and unsuper-vised target domain data. The FM encoder adapter and decoder are then finetuned to the target domain with a small amount of super-vised in-domain data. On a large-scale YouTube and Voice Search task, our method is shown to be both data and model parameter efficient. It achieves the same quality with only 21.6M supervised in-domain data and 130.8M finetuned parameters, compared to the 731.1M model trained from scratch on additional 300M supervised in-domain data.

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.

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

In this work, we present a model-based approach to improving contextual biasing that improves quality without drastically increasing model computation during inference. Specifically, we look at injecting text data during training which is representative of contextually-relevant context that will be seen at inference, using a modality-matching text injection method known as JOIST. As JOIST injects text data directly into the E2E model, there is no additional model computation during inference, which is a big difference compared to most model-based biasing techniques. We find that our proposed approach, when combined with an FST-based context model, improves recognition of contacts between 5–15% relative.

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

We propose a quantum kernel learning (QKL) framework to address the inherent data sparsity issues often encountered in training large-scare acoustic models in low-resource scenarios. We project acoustic features based on classical-to-quantum feature encoding. Different from existing quantum convolution techniques, we utilize QKL with features in the quantum space to design kernel-based classifiers. Experimental results on challenging spoken command recognition tasks for a few low-resource languages, such as Arabic, Georgian, Chuvash, and Lithuanian, show that the proposed QKL-based hybrid approach attains good improvements over existing classical and quantum solutions.

Word-piece models (WPMs) are commonly used subword units in state-of-the-art end-to-end automatic speech recognition (ASR) systems. For multilingual ASR, due to the differences in written scripts across languages, multilingual WPMs bring the challenges of having overly large output layers and scaling to more languages. In this work, we propose a universal monolingual output layer (UML) to address such problems. Instead of one output node for only one WPM, UML re-associates each output node with multiple WPMs, one for each language, and results in a smaller monolingual output layer shared across languages. Consequently, the UML enables to switch in the interpretation of each output node depending on the language of the input speech. Experimental results on an 11-language voice search task demonstrated the feasibility of using UML for high-quality and high-efficiency multilingual streaming ASR.

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

Self-supervised training has shown promising gains in pretraining models and facilitating the downstream finetuning for speech recognition, like multilingual ASR. Most existing methods adopt a 2-stage scheme where the self-supervised loss is optimized in the first pretraining stage, and the standard supervised finetuning resumes in the second stage. In this paper, we propose an end-to-end (E2E) Joint Unsupervised and Supervised Training (JUST) method to combine the supervised RNN-T loss and the self-supervised contrastive and masked language modeling (MLM) losses. We validate its performance on the public dataset Multilingual LibriSpeech (MLS), which includes 8 languages and is extremely imbalanced. On MLS, we explore (1) JUST trained from scratch, and (2) JUST finetuned from a pretrained checkpoint. Experiments show that JUST can consistently outperform other existing state-of-the-art methods, and beat the monolingual baseline by a significant margin, demonstrating JUST’s capability of handling low-resource languages in multilingual ASR. Our average WER of all languages outperforms average monolingual baseline by 33.3%, and the state-of-the-art 2-stage XLSR by 32%. On low-resource languages like Polish, our WER is less than half of the monolingual baseline and even beats the supervised transfer learning method which uses external supervision.

Previous research on applying deliberation networks to automatic speech recognition has achieved excellent results. The attention decoder based deliberation model often works as a rescorer to improve first-pass recognition results, and requires the full first-pass hypothesis for second-pass deliberation. In this work, we propose a transducer-based streaming deliberation model. The joint network of a transducer decoder often receives inputs from the encoder and the prediction network. We propose to use attention to the first-pass text hypothesis as the third input to the joint network. The proposed transducer based deliberation model naturally streams, making it more desirable for on-device applications. We also show that the model improves rare word recognition compared to cascaded encoders, with relative WER reductions ranging from 3.6% to 10.4% for a variety of test sets. Our model does not use any additional text data for training.

The development of end-to-end models has largely sped up the research in massively multilingual automatic speech recognition (MMASR). Previous research has demonstrated the feasibility to build high quality MMASR models. In this work, we study the impact of adding more languages and propose a lifelong learning approach to build high quality MMASR systems. Experiments on a 66-language Voice Search task show that we can take a model built on 15 languages and continue training to obtain a 32-language model and similarly to further build a 67-language model. More importantly, models developed in this way achieve better quality compared to those trained from scratch. It maintains similar performance on old languages and achieves competitive results on new ones. This would potentially speed up the development of universal ASR models that recognize speech from any language, any domain and any environment by reusing knowledge learned beforehand.

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.

We propose to deliberate the hypothesis alignment of a streaming RNN-T model with the previously proposed Align-Refine non-autoregressive decoding method and its improved versions. The method performs a few refinement steps, where each step shares a transformer decoder that attends to both text features (extracted from alignments) and audio features, and outputs complete updated alignments. The transformer decoder is trained with the CTC loss which facilitates parallel greedy decoding, and performs full-context attention to capture label dependencies. We improve Align-Refine by introducing cascaded encoder that captures more audio context before refinement, and alignment augmentation which enforces learning label dependency. We show that, conditioned on hypothesis alignments of a streaming RNN-T model, our method obtains significantly more accurate recognition results than the first-pass RNN-T, with only small amount of model parameters.

The recurrent neural network transducer (RNN-T) has recently become the mainstream end-to-end approach for streaming automatic speech recognition (ASR). To estimate the output distributions over subword units, RNN-T uses a fully connected layer as the joint network to fuse the acoustic representations extracted using the acoustic encoder with the text representations obtained using the prediction network based on the previous subword units. In this paper, we propose to use gating, bilinear pooling, and a combination of them in the joint network to produce more expressive representations to feed into the output layer. A regularisation method is also proposed to enable better acoustic encoder training by reducing the gradients back-propagated into the prediction network at the beginning of RNN-T training. Experimental results on a multilingual ASR setting for voice search over nine languages show that the joint use of the proposed methods can result in 4%–5% relative word error rate reductions with only a few million extra parameters.

In voice-enabled applications, a predetermined hotword is usually used to activate a device in order to attend to the query.However, speaking queries followed by a hotword each time introduces a cognitive burden in continued conversations.To avoid repeating a hotword, we propose a streaming end-to-end (E2E) intended query detector that identifies the utterances directed towards the device and filters out other utterances not directed towards device.The proposed approach incorporates the intended query detector into the E2E model that already folds different components of the speech recognition pipeline into one neural network.The E2E modeling on speech decoding and intended query detection also allows us to declare a quick intended query detection based on early partial recognition result, which is important to decrease latency and make the system responsive.We demonstrate that the proposed E2E approach yields a 22% relative improvement on equal error rate (EER) for the detection accuracy and 600 ms latency improvement compared with an independent intended query detector.In our experiment, the proposed model detects whether the user is talking to the device with a 8.7% EER within 1.4 seconds of median latency after user starts speaking.

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.

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

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

On-device end-to-end (E2E) models have shown improvements over a conventional model on English Voice Search tasks in both quality and latency.E2E models have also shown promising results for multilingual automatic speech recognition (ASR).In this paper, we extend our previous capacity solution to streaming applications and present a streaming multilingual E2E ASR system that runs fully on device with comparable quality and latency to individual monolingual models.To achieve that, we propose an Encoder Endpointer model and an End-of-Utterance (EOU) Joint Layer for a better quality and latency trade-off.Our system is built in a language agnostic manner allowing it to natively support intersentential code switching in real time.To address the feasibility concerns on large models, we conducted on-device profiling and replaced the time consuming LSTM decoder with the recently developed Embedding decoder.With these changes, we managed to run such a system on a mobile device in less than real time.

The careful construction of audio representations has become a dominant feature in the design of approaches to many speech tasks.Increasingly, such approaches have emphasized "disentanglement", where a representation contains only parts of the speech signal relevant to transcription while discarding irrelevant information.In this paper, we construct a representation learning task based on joint modeling of ASR and TTS, and seek to learn a representation of audio that disentangles that part of the speech signal that is relevant to transcription from that part which is not.We present empirical evidence that successfully finding such a representation is tied to the randomness inherent in training.We then make the observation that these desired, disentangled solutions to the optimization problem possess unique statistical properties.Finally, we show that enforcing these properties during training improves WER by 24.5% relative on average for our joint modeling task.These observations motivate a novel approach to learning effective audio representations.

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.

We propose a streaming non-autoregressive (non-AR) decoding algorithm to deliberate the hypothesis alignment of a streaming RNN-T model.Our algorithm facilitates a simple greedy decoding procedure, and at the same time is capable of producing the decoding result at each frame with limited right context, thus enjoying both high efficiency and low latency.These advantages are achieved by converting the offline Align-Refine algorithm to be streaming-compatible, with a novel transformer decoder architecture that performs local self-attentions for both text and audio, and a time-aligned cross-attention at each layer.Furthermore, we perform discriminative training of our model with the minimum word error rate (MWER) criterion, which has not been done in the non-AR decoding literature.Experiments on voice search datasets and Librispeech show that with reasonable right context, our streaming model performs as well as the offline counterpart, and discriminative training leads to further WER gain when the first-pass model has small capacity.

Automatic speech recognition (ASR) systems typically rely on an external endpointer (EP) model to identify speech boundaries. In this work, 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 results in a single E2E model that can be used during inference to perform frame filtering at low cost, and also make high quality end-of-query (EOQ) predictions based on ongoing ASR computation. We present results on a voice search test set showing that, compared to separate single-task models, this approach reduces median endpoint latency by 120 ms (30.8% reduction), and 90th percentile latency by 170 ms (23.0% reduction), without regressing word error rate. For continuous recognition, WER improves by 10.6% (relative).

Multilingual end-to-end automatic speech recognition models are attractive due to its simplicity in training and deployment. Recent work on large-scale training of such models has shown promising results compared to monolingual models. However, the work often focuses on multilingual models themselves in a single-pass setup. In this work, we investigate second-pass deliberation for multilingual speech recognition. Our proposed deliberation is multilingual, i.e., the text encoder encodes hypothesis text from multiple languages, and the decoder attends to multilingual text and audio. We investigate scaling the deliberation text encoder and decoder, and compare scaling the deliberation decoder and the first-pass cascaded encoder. We show that deliberation improves the average WER on 9 languages by 4% relative compared to the single-pass model. By increasing the size of the deliberation up to 1B parameters, the average WER improvement increases to 9%, with up to 14% for certain languages. Our deliberation rescorer is based on transformer layers and can be parallelized during rescoring.

Automatic speech recognition (ASR) systems need to be accurate, have low latency, and effectively handle language switching in order to be useful for the 60% of the world population that speaks more than one language. Thus, we propose a truly multilingual first-pass and monolingual second-pass streaming on-device ASR system based on the recently developed Cascaded Encoders model. The streaming first-pass recognizes multilingual speech without needing language information, providing real-time transcription, even for code-switching speech. The second-pass uses a language dependent right context encoder to improve the recognition accuracy. On a 9 language Voice Search task, we find that a system combining shared causal encoder with decoders and non-causal encoders replicated per-language reduces word error rate (WER) by 4.4% relative to monolingual baselines. We further show this design to be parameter efficient, outperforming other architectures when matched in the number of parameters.

Attention-based biasing techniques for end-to-end ASR systems are able to achieve large accuracy gains without requiring the inference algorithm adjustments and parameter tuning common to fusion approaches. However, it is challenging to simultaneously scale up attention-based biasing to realistic numbers of biased phrases; maintain in-domain WER gains, while minimizing out-of-domain losses; and run in real time. We present NAM+, an attention-based biasing approach which achieves a 16X inference speedup per acoustic frame over prior work when run with 3,000 biasing entities, as measured on a typical mobile CPU. NAM+ achieves these run-time gains through a combination of Two-Pass Hierarchical Attention and Dilated Context Update. Compared to the adapted baseline, NAM+ further decreases the in-domain WER by up to 12.6% relative, while incurring an out-of-domain WER regression of 20% relative. Compared to the non-adapted baseline, the out-of-domain WER regression is 7.1 % relative.

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

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

Building ASR models across many languages is a challenging multi-task learning problem due to large variations and heavily unbalanced data. Existing work has shown positive transfer from high resource to low resource languages. However, degradations on high resource languages are commonly observed due to interference from the heterogeneous multilingual data and reduction in per-language capacity. We conduct a capacity study on a 15-language task, with the amount of data per language varying from 7.6K to 53.5K hours. We adopt GShard [1] to efficiently scale up to 10B parameters. Empirically, we find that (1) scaling the number of model parameters is an effective way to solve the capacity bottleneck - our 500M-param model already outperforms monolingual baselines and scaling it to 1B and 10B brought further quality gains; (2) larger models are not only more data efficient, but also more efficient in terms of training cost as measured in TPU days - the 1B-param model reaches the same accuracy at 34% of training time as the 500M-param model; (3) given a fixed capacity budget, adding depth works better than width and large encoders do better than large decoders; (4) with continuous training, they can be adapted to new languages and domains.

End-to-end (E2E) models have shown to outperform state-of-the-art conventional models for streaming speech recognition [1] across many dimensions, including quality (as measured by word error rate (WER)) and endpointer latency [2]. However, the model still tends to delay the predictions towards the end and thus has much higher partial latency compared to a conventional ASR model. To address this issue, we look at encouraging the E2E model to emit words early, through an algorithm called FastEmit [3]. Naturally, improving on latency results in a quality degradation. To address this, we explore replacing the LSTM layers in the encoder of our E2E model with Conformer layers [4], which has shown good improvements for ASR. Secondly, we also explore running a 2nd-pass beam search to improve quality. In order to ensure the 2nd-pass completes quickly, we explore non-causal Conformer layers that feed into the same 1st-pass RNN-T decoder, an algorithm called Cascaded Encoders [5]. Overall, the Conformer RNN-T with Cascaded Encoders offers a better quality and latency tradeoff for streaming ASR.

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

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.

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.

We propose automatic speech recognition (ASR) models inspired by echo state network (ESN) [1], in which a subset of recurrent neural networks (RNN) layers in the models are randomly initialized and untrained. Our study focuses on RNN-T and Conformer models, and we show that model quality does not drop even when the decoder is fully randomized. Furthermore, such models can be trained more efficiently as the decoders do not require to be updated. By contrast, randomizing encoders hurts model quality, indicating that optimizing encoders and learn proper representations for acoustic inputs are more vital for speech recognition. Overall, we challenge the common practice of training ASR models for all components, and demonstrate that ESN-based models can perform equally well but enable more efficient training and storage than fully-trainable counterparts.

Streaming automatic speech recognition (ASR) aims to emit each hypothesized word as quickly and accurately as possible. However, emitting fast without degrading quality, as measured by word error rate (WER), is highly challenging. Existing approaches including Early and Late Penalties [1] and Constrained Alignments [2], [3] penalize emission delay by manipulating per-token or per-frame probability prediction in sequence transducer models [4]. While being successful in reducing delay, these approaches suffer from significant accuracy regression and also require additional word alignment information from an existing model. In this work, we propose a sequence-level emission regularization method, named FastEmit, that applies latency regularization directly on per-sequence probability in training transducer models, and does not require any alignment. We demonstrate that FastEmit is more suitable to the sequence-level optimization of transducer models [4] for streaming ASR by applying it on various end-to-end streaming ASR networks including RNN-Transducer [5], Transformer-Transducer [6], [7], ConvNet-Transducer [8] and Conformer-Transducer [9]. We achieve 150 ~ 300ms latency reduction with significantly better accuracy over previous techniques 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 210ms to only 30ms on LibriSpeech.

Previous works on the Recurrent Neural Network-Transducer (RNN-T) models have shown that, under some conditions, it is possible to simplify its prediction network with little or no loss in recognition accuracy (arXiv:2003.07705 [eess.AS], [2], arXiv:2012.06749 [cs.CL]). This is done by limiting the context size of previous labels and/or using a simpler architecture for its layers instead of LSTMs. The benefits of such changes include reduction in model size, faster inference and power savings, which are all useful for on-device applications. In this work, we study ways to make the RNN-T decoder (prediction network + joint network) smaller and faster without degradation in recognition performance. Our prediction network performs a simple weighted averaging of the input embeddings, and shares its embedding matrix weights with the joint network's output layer (a.k.a. weight tying, commonly used in language modeling arXiv:1611.01462 [cs.LG]). This simple design, when used in conjunction with additional Edit-based Minimum Bayes Risk (EMBR) training, reduces the RNN-T Decoder from 23M parameters to just 2M, without affecting word-error rate (WER).

We introduce Lookup-Table Language Models (LookupLM), a method for scaling up the size of RNN language models with only a constant increase in the floating point operations, by increasing the expressivity of the embedding table. In particular, we instantiate an (additional) embedding table which embeds the previous n-gram token sequence, rather than a single token. This allows the embedding table to be scaled up arbitrarily -- with a commensurate increase in performance -- without changing the token vocabulary. Since embeddings are sparsely retrieved from the table via a lookup; increasing the size of the table adds neither extra operations to each forward pass nor extra parameters that need to be stored on limited GPU/TPU memory. We explore scaling n-gram embedding tables up to nearly a billion parameters. When trained on a 3-billion sentence corpus, we find that LookupLM improves long tail log perplexity by 2.44 and long tail WER by 23.4% on a downstream speech recognition task over a standard RNN language model baseline, an improvement comparable to a scaling up the baseline by 6.2x the number of floating point operations.

We propose a new two-pass E2E speech recognition model that improves ASR performance by training on a combination of paired data and unpaired text data.Previously, the joint acoustic and text decoder (JATD) has shown promising results through the use of text data during model training and the recently introduced deliberation architecture has reduced recognition errors by leveraging first-pass decoding results.Our method, dubbed Deliberation-JATD, combines the spelling correcting abilities of deliberation with JATD's use of unpaired text data to further improve performance.The proposed model produces substantial gains across multiple test sets, especially those focused on rare words, where it reduces word error rate (WER) by between 12% and 22.5% relative.This is done without increasing model size or requiring multi-stage training, making Deliberation-JATD an efficient candidate for on-device applications.

We propose a multitask training method for attention-based end-to-end speech recognition models.We regularize the decoder in a listen, attend, and spell model by multitask training it on both audio-text and text-only data.Trained on the 100-hour subset of LibriSpeech, the proposed method, without requiring an additional language model, leads to an 11% relative performance improvement over the baseline and approaches the performance of language model shallow fusion on the test-clean evaluation set.We observe a similar trend on the whole 960hour LibriSpeech training set.Analyses of different types of errors and sample output sentences demonstrate that the proposed method can incorporate language level information, suggesting its effectiveness in real-world applications.

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

Interactive speech recognition systems must generate words quickly while also producing accurate results. Two-pass models excel at these requirements by employing a first-pass decoder that quickly emits words, and a second-pass decoder that requires more context but is more accurate. Previous work has established that a deliberation network can be an effective second-pass model. The model attends to two kinds of inputs at once: encoded audio frames and the hypothesis text from the first-pass model. In this work, we explore using transformer layers instead of long-short term memory (LSTM) layers for deliberation rescoring. In transformer layers, we generalize the "encoder-decoder" attention to attend to both encoded audio and first-pass text hypotheses. The output context vectors are then combined by a merger layer. Compared to LSTM-based deliberation, our best transformer deliberation achieves 7% relative word error rate improvements along with a 38% reduction in computation. We also compare against non-deliberation transformer rescoring, and find a 9% relative improvement.

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

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 has shown promising potential for on-device ASR [1]. 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 [2] with additional early and late penalties. By further applying the minimum word error rate (MWER) training technique [3], we achieved 8.0% relative word error rate (WER) reduction and 130ms 90-percentile latency reduction over [2] on a Voice Search test set. We also experimented with a second-pass Listen, Attend and Spell (LAS) rescorer [4]. Although it did not directly improve the first pass latency, the large WER reduction provides extra room to trade WER for latency. RNN-T EP+LAS, together with MWER training brings in 18.7% relative WER reduction and 160ms 90-percentile latency reductions compared to the original proposed RNN-T EP [2] model.

Proper nouns present a challenge for end-to-end (E2E) automatic speech recognition (ASR) systems in that a particular name may appear only rarely during training, and may have a pronunciation similar to that of a more common word. Unlike conventional ASR models, E2E systems lack an explicit pronounciation model that can be specifically trained with proper noun pronounciations and a language model that can be trained on a large text-only corpus. Past work has addressed this issue by incorporating additional training data or additional models. In this paper, we instead build on recent advances in minimum word error rate (MWER) training to develop two new loss criteria that specifically emphasize proper noun recognition. Unlike past work on this problem, this method requires no new data during training or external models during inference. We see improvements ranging from 2% to 7% relative on several relevant benchmarks.

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.

Recently, we introduced a two-pass on-device end-to-end (E2E) speech recognition model, which runs RNN-T in the first-pass and then rescores/redecodes the result using a noncausal Listen, Attend and Spell (LAS) decoder. This on-device model obtained similar performance to a state-of-the-art conventional model. However, like many E2E models, it suffers from being trained only on supervised audio-text pairs and thus performs poorly on rare words compared to a conventional model which incorporates a language model trained on a much larger text corpus. In this work, we introduce a joint acoustic and text decoder (JATD) into the LAS decoder, which makes it possible to incorporate a much larger text corpus into training. We find that the JATD model obtains in a 3-10% relative improvement in WER compared to a LAS decoder trained only on supervised audio-text pairs across a variety of proper noun test sets.

Recurrent Neural Network Transducer (RNN-T) models [1] for automatic speech recognition (ASR) provide high accuracy speech recognition. Such end-to-end (E2E) models combine acoustic, pronunciation and language models (AM, PM, LM) of a conventional ASR system into a single neural network, dramatically reducing complexity and model size.In this paper, we propose a technique for incorporating contextual signals, such as intelligent assistant device state or dialog state, directly into RNN-T models. We explore different encoding methods and demonstrate that RNN-T models can effectively utilize such context. Our technique results in reduction in Word Error Rate (WER) of up to 10.4% relative on a variety of contextual recognition tasks. We also demonstrate that proper regularization can be used to model context independently for improved overall quality.

End-to-end (E2E) automatic speech recognition (ASR) systems lack the distinct language model (LM) component that characterizes traditional speech systems. While this simplifies the model architecture, it complicates the task of incorporating text-only data into training, which is important to the recognition of tail words that do not occur often in audio-text pairs. While shallow fusion has been proposed as a method for incorporating a pre-trained LM into an E2E model at inference time, it has not yet been explored for very large text corpora, and it has been shown to be very sensitive to hyperparameter settings in the beam search. In this work, we apply shallow fusion to incorporate a very large text corpus into a state-of-the-art E2EASR model. We explore the impact of model size and show that intelligent pruning of the training set can be more effective than increasing the parameter count. Additionally, we show that incorporating the LM in minimum word error rate (MWER) fine tuning makes shallow fusion far less dependent on optimal hyperparameter settings, reducing the difficulty of that tuning problem.

End-to-end automatic speech recognition (ASR) models, including both attention-based models and the recurrent neural network transducer (RNN-T), have shown superior performance compared to conventional systems [1], [2]. However, previous studies have focused primarily on short utterances that typically last for just a few seconds or, at most, a few tens of seconds. Whether such architectures are practical on long utterances that last from minutes to hours remains an open question. In this paper, we both investigate and improve the performance of end-to-end models on long-form transcription. We first present an empirical comparison of different end-to-end models on a real world long-form task and demonstrate that the RNN-T model is much more robust than attention-based systems in this regime. We next explore two improvements to attention-based systems that significantly improve its performance: restricting the attention to be monotonic, and applying a novel decoding algorithm that breaks long utterances into shorter overlapping segments. Combining these two improvements, we show that attention-based end-to-end models can be very competitive to RNN-T on long-form speech recognition.

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

Improving the representation of contextual information is key to unlocking the potential of end-to-end (E2E) automatic speech recognition (ASR). In this work, we present a novel and simple approach for training an ASR context mechanism with difficult negative examples. The main idea is to focus on proper nouns (e.g., unique entities such as names of people and places) in the reference transcript and use phonetically similar phrases as negative examples, encouraging the neural model to learn more discriminative representations. We apply our approach to an end-to-end contextual ASR model that jointly learns to transcribe and select the correct context items. We show that our proposed method gives up to 53.1% relative improvement in word error rate (WER) across several benchmarks.

End-to-end (E2E) models are a promising research direction in speech recognition, as the single all-neural E2E system offers a much simpler and more compact solution compared to a conventional model, which has a separate acoustic (AM), pronunciation (PM) and language model (LM). However, it has been noted that E2E models perform poorly on tail words and proper nouns, likely because the end-to-end optimization requires joint audio-text pairs, and does not take advantage of additional lexicons and large amounts of text-only data used to train the LMs in conventional models. There has been numerous efforts in training an RNN-LM on text-only data and fusing it into the end-to-end model. In this work, we contrast this approach to training the E2E model with audio-text pairs generated from unsupervised speech data. To target the proper noun issue specifically, we adopt a Part-of-Speech (POS) tagger to filter the unsupervised data to use only those with proper nouns. We show that training with filtered unsupervised-data provides up to a 13% relative reduction in word-error-rate (WER), and when used in conjunction with a cold-fusion RNN-LM, up to a 17% relative improvement.

End-to-End (E2E) automatic speech recognition (ASR) systems learn word spellings directly from text-audio pairs, in contrast to traditional ASR systems which incorporate a separate pronunciation lexicon. The lexicon allows a traditional system to correctly spell rare words observed only in LM training, if their phonetic pronunciation is known during inference. E2E systems, however, are more likely to misspell rare words.We propose an E2E model which benefits from the best of both worlds: it outputs graphemes, and thus learns to spell words directly, while leveraging pronunciations for words which might be likely in a given context. Our model is based on the recently proposed Contextual Listen, Attend, and Spell (CLAS) model. As in CLAS, our model accepts a set of bias phrases, which are first converted into fixed length embeddings which are provided as additional inputs to the model. Unlike CLAS, which accepts only the textual form of the bias phrases, the proposed model also has access to the corresponding phonetic pronunciations, which improves performance on challenging sets which include words unseen in training. The proposed model provides a 16% relative word-error-rate reduction over CLAS when both the phonetic and written representation of the context bias phrases are used.

The tradeoff between word error rate (WER) and latency is very important for streaming automatic speech recognition (ASR) applications. We want the system to endpoint and close the microphone as quickly as possible, without degrading WER. Conventional ASR systems rely on a separately trained endpointing module, which interacts with the acoustic, pronunciation and language model (AM, PM, and LM) components, and can result in a higher WER or a larger latency. In going with the all-neural spirit of end-to-end (E2E) models, which fold the AM, PM and LM into a single neural network, in this work we look at folding the endpointer into this E2E model to assist with the endpointing task. We refer to this jointly optimized model - which performs both recognition and endpointing - as an E2E enpointer. On a large vocabulary Voice Search task, we show that the combination of such an E2E endpoiner with a conventional endpointer results in no quality degradation, while reducing latency by more than a factor of 2 compared to using a separate endpointer with the E2E model.

Attention-based sequence-to-sequence models for speech recognition jointly train an acoustic model, language model (LM), and alignment mechanism using a single neural network and require only parallel audio-text pairs. Thus, the language model component of the end-to-end model is only trained on transcribed audio-text pairs, which leads to performance degradation especially on rare words. While there have been a variety of work that look at incorporating an external LM trained on text-only data into the end-to-end framework, none of them have taken into account the characteristic error distribution made by the model. In this paper, we propose a novel approach to utilizing text-only data, by training a spelling correction (SC) model to explicitly correct those errors. On the LibriSpeech dataset, we demonstrate that the proposed model results in an 18.6% relative improvement in WER over the baseline model when directly correcting top ASR hypothesis, and a 29.0% relative improvement when further rescoring an expanded n-best list using an external LM.

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.

We present two end-to-end models: Audio-to-Byte (A2B) and Byte-to-Audio (B2A), for multilingual speech recognition and synthesis. Prior work has predominantly used characters, sub-words or words as the unit of choice to model text. These units are difficult to scale to languages with large vocabularies, particularly in the case of multilingual processing. In this work, we model text via a sequence of Unicode bytes, specifically, the UTF-8 variable length byte sequence for each character. Bytes allow us to avoid large softmaxes in languages with large vocabularies, and share representations in multilingual models. We show that bytes are superior to grapheme characters over a wide variety of languages in monolingual end-to-end speech recognition. Additionally, our multilingual byte model outperform each respective single language baseline on average by 4.4% relatively. In Japanese-English code-switching speech, our multilingual byte model outperform our monolingual baseline by 38.6% relatively. Finally, we present an end-to-end multilingual speech synthesis model using byte representations which matches the performance of our monolingual baselines.

A method (500) includes receiving audio data encoding an utterance (106) spoken by a native speaker (110) of a first language, and receiving a biasing term list (105) including one or more terms in a second language different than the first language. The method also includes processing, using a speech recognition model (200), acoustic features (105) derived from the audio data to generate speech recognition scores for both wordpieces and corresponding phoneme sequences in the first language. The method also includes rescoring the speech recognition scores for the phoneme sequences based on the one or more terms in the biasing term list, and executing, using the speech recognition scores for the wordpieces and the rescored speech recognition scores for the phoneme sequences, a decoding graph (400) to generate a transcription (116) for the utterance.

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).

Recognizing written domain numeric utterances (e.g., I need $1.25.) can be challenging for ASR systems, particularly when numeric sequences are not seen during training.This out-ofvocabulary (OOV) issue is addressed in conventional ASR systems by training part of the model on spoken domain utterances (e.g., I need one dollar and twenty five cents.),for which numeric sequences are composed of in-vocabulary numbers, and then using an FST verbalizer to denormalize the result.Unfortunately, conventional ASR models are not suitable for the low memory setting of on-device speech recognition.E2E models such as RNN-T are attractive for on-device ASR, as they fold the AM, PM and LM of a conventional model into one neural network.However, in the on-device setting the large memory footprint of an FST denormer makes spoken domain training more difficult.In this paper, we investigate techniques to improve E2E model performance on numeric data.We find that using a text-to-speech system to generate additional numeric training data, as well as using a small-footprint neural network to perform spoken-to-written domain denorming, yields improvement in several numeric classes.In the case of the longest numeric sequences, we see reduction of WER by up to a factor of 8.

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.

Voice activity detection (VAD) is the task of predicting which parts of an utterance contains speech versus background noise. It is an important first step to determine which samples to send to the decoder and when to close the microphone. The long short-term memory neural network (LSTM) is a popular architecture for sequential modeling of acoustic signals, and has been successfully used in several VAD applications. However, it has been observed that LSTMs suffer from state saturation problems when the utterance is long (i.e., for voice dictation tasks), and thus requires the LSTM state to be periodically reset. In this paper, we propose an alternative architecture that does not suffer from saturation problems by modeling temporal variations through a stateless dilated convolution neural network (CNN). The proposed architecture differs from conventional CNNs in three respects: it uses dilated causal convolution, gated activations and residual connections. Results on a Google Voice Typing task shows that the proposed architecture achieves 14% relative FA improvement at a FR of 1% over state-of-the-art LSTMs for VAD task. We also include detailed experiments investigating the factors that distinguish the proposed architecture from conventional convolution.

Attention-based encoder-decoder architectures such as Listen, Attend, and Spell (LAS), subsume the acoustic, pronunciation and language model components of a traditional automatic speech recognition (ASR) system into a single neural network. In previous work, we have shown that such architectures are comparable to state-of-the-art ASR systems on dictation tasks, but it was not clear if such architectures would be practical for more challenging tasks such as voice search. In this work, we explore a variety of structural and optimization improvements to our LAS model which significantly improve performance. On the structural side, we show that word piece models can be used instead of graphemes. We also introduce a multi-head attention architecture, which offers improvements over the commonly-used single-head attention. On the optimization side, we explore synchronous training, scheduled sampling, label smoothing, and minimum word error rate optimization, which are all shown to improve accuracy. We present results with a unidirectional LSTM encoder for streaming recognition. On a 12, 500 hour voice search task, we find that the proposed changes improve the WER from 9.2% to 5.6%, while the best conventional system achieves 6.7%; on a dictation task our model achieves a WER of 4.1% compared to 5% for the conventional system.

Mask based statistical beamforming, where signal statistics for the target and the interference gained from masking are used for beamforming, has shown great effectiveness in the two recent CHiME challenges. This idea has sparked interest in the research community and resulted in numerous proposed approaches based on the idea. At the same time, the advent of voice controlled smart home devices, such as Google Home and Amazon Alexa, has strengthened the need for robust far-field automatic speech recognition. In this paper, we evaluate if mask based beamforming can live up to the expectations created by the CHiME challenges and provide similar gains in a smart home scenario. To this extend, we pinpoint the main differences between the scenarios, review the recent developments and conduct extensive experiments on large scale data. These experiments show that, while a 10 % relative reduction of the word error rate can be achieved, the gains are not as high as those seen in the CHiME challenge. We also show that approaches where the frontend and back-end is trained jointly do not reach the performance level of their independently trained counterparts. On the plus side, we see a 20 % relative improvement for an evaluation set with crosstalk.

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

In this paper, we present an algorithm which introduces phase-perturbation to the training database when training phase-sensitive deep neural-network models. Traditional features such as log-mel or cepstral features do not have have any phase-relevant information. However features such as raw-waveform or complex spectra features contain phase-relevant information. Phase-sensitive features have the advantage of being able to detect differences in time of arrival across different microphone channels or frequency bands. However, compared to magnitude-based features, phase information is more sensitive to various kinds of distortions such as variations in microphone characteristics, reverberation, and so on. For traditional magnitude-based features, it is widely known that adding noise or reverberation, often called Multistyle-TRaining (MTR), improves robustness. In a similar spirit, we propose an algorithm which introduces spectral distortion to make the deep-learning models more robust to phase-distortion. We call this approach Spectral-Distortion TRaining (SDTR). In our experiments using a training set consisting of 22-million utterances with and without MTR, this approach reduces Word Error Rates (WERs) relatively by 3.2 % and 8.48 % respectively on test sets recorded on Google Home.

Sequence-to-sequence models provide a simple and elegant solution for building speech recognition systems by folding separate components of a typical system, namely acoustic (AM), pronunciation (PM) and language (LM) models into a single neural network. In this work, we look at one such sequence-to-sequence model, namely listen, attend and spell (LAS) [1], and explore the possibility of training a single model to serve different English dialects, which simplifies the process of training multi-dialect systems without the need for separate AM, PM and LMs for each dialect. We show that simply pooling the data from all dialects into one LAS model falls behind the performance of a model fine-tuned on each dialect. We then look at incorporating dialect-specific information into the model, both by modifying the training targets by inserting the dialect symbol at the end of the original grapheme sequence and also feeding a 1-hot representation of the dialect information into all layers of the model. Experimental results on seven English dialects show that our proposed system is effective in modeling dialect variations within a single LAS model, outperforming a LAS model trained individually on each of the seven dialects by 3.1~16.5% relative.

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

Having a sequence-to-sequence model which can operate in an online fashion is important for streaming applications such as Voice Search. Neural transducer is a streaming sequence-to-sequence model, but has shown a significant degradation in performance compared to non-streaming models such as Listen, Attend and Spell (LAS). In this paper, we present various improvements to NT. Specifically, we look at increasing the window over which NT computes attention, mainly by looking backwards in time so the model still remains online. In addition, we explore initializing a NT model from a LAS-trained model so that it is guided with a better alignment. Finally, we explore including stronger language models such as using wordpiece models, and applying an external LM during the beam search. On a Voice Search task, we find with these improvements we can get NT to match the performance of LAS.

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

Training a conventional automatic speech recognition (ASR) system to support multiple languages is challenging because the sub-word unit, lexicon and word inventories are typically language specific. In contrast, sequence-to-sequence models are well suited for multilingual ASR because they encapsulate an acoustic, pronunciation and language model jointly in a single network. In this work we present a single sequence-to-sequence ASR model trained on 9 different Indian languages, which have very little overlap in their scripts. Specifically, we take a union of language-specific grapheme sets and train a grapheme-based sequence-to-sequence model jointly on data from all languages. We find that this model, which is not explicitly given any information about language identity, improves recognition performance by 21% relative compared to analogous sequence-to-sequence models trained on each language individually. By modifying the model to accept a language identifier as an additional input feature, we further improve performance by an additional 7% relative and eliminate confusion between different languages.

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

Attention-based recurrent neural encoder-decoder models present an elegant solution to the automatic speech recognition problem. This approach folds the acoustic model, pronunciation model, and language model into a single network and requires only a parallel corpus of speech and text for training. However, unlike in conventional approaches that combine separate acoustic and language models, it is not clear how to use additional (unpaired) text. While there has been previous work on methods addressing this problem, a thorough comparison among methods is still lacking. In this paper, we compare a suite of past methods and some of our own proposed methods for using unpaired text data to improve encoder-decoder models. For evaluation, we use the medium-sized Switchboard data set and the large-scale Google voice search and dictation data sets. Our results confirm the benefits of using unpaired text across a range of methods and data sets. Surprisingly, for first-pass decoding, the rather simple approach of shallow fusion performs best across data sets. However, for Google data sets we find that cold fusion has a lower oracle error rate and outperforms other approaches after second-pass rescoring on the Google voice search data set.

Sequence-level losses are commonly used to train deep neural network acoustic models for automatic speech recognition. The forward-backward algorithm is used to efficiently compute the gradients of the sequence loss with respect to the model parameters. Gradient-based optimization is used to minimize these losses. Recent work has shown that the forward-backward algorithm can be efficiently implemented as a series of matrix operations. This paper further improves the forward-backward algorithm via batched computation, a technique commonly used to improve training speed by exploiting the parallel computation of matrix multiplication. Specifically, we show how batched computation of the forward-backward algorithm can be efficiently implemented using TensorFlow to handle variable-length sequences within a mini batch. Furthermore, we also show how the batched forward-backward computation can be used to compute the gradients of the connectionist temporal classification (CTC) and maximum mutual information (MMI) losses with respect to the logits. We show, via empirical benchmarks, that the batched forward-backward computation can speed up the CTC loss and gradient computation by about 183 times when run on GPU with a batch size of 256 compared to using a batch size of 1; and by about 22 times for lattice-free MMI using a trigram phone language model for the denominator.

Multichannel automatic speech recognition (ASR) systems commonly separate speech enhancement, including localization, beamforming, and postfiltering, from acoustic modeling. In this paper, we perform multichannel enhancement jointly with acoustic modeling in a deep neural network framework. Inspired by beamforming, which leverages differences in the fine time structure of the signal at different microphones to filter energy arriving from different directions, we explore modeling the raw time-domain waveform directly. We introduce a neural network architecture, which performs multichannel filtering in the first layer of the network, and show that this network learns to be robust to varying target speaker direction of arrival, performing as well as a model that is given oracle knowledge of the true target speaker direction. Next, we show how performance can be improved by factoring the first layer to separate the multichannel spatial filtering operation from a single channel filterbank which computes a frequency decomposition. We also introduce an adaptive variant, which updates the spatial filter coefficients at each time frame based on the previous inputs. Finally, we demonstrate that these approaches can be implemented more efficiently in the frequency domain. Overall, we find that such multichannel neural networks give a relative word error rate improvement of more than 5% compared to a traditional beamforming-based multichannel ASR system and more than 10% compared to a single channel waveform model.

Deep Neural Networks (DNNs) have recently been shown to significantly outperform existing machine learning techniques in several pattern recognition tasks. DNNs are the state-of-the-art models used in image recognition, object detection, classification and tracking, and speech and language processing applications. The biggest drawback to DNNs has been the enormous cost in computation and time taken to train the parameters of the networks-often a tenfold increase relative to conventional technologies. Such training time costs can be mitigated by the application of parallel computing algorithms and architectures. However, these algorithms often run into difficulties because of the cost of inter-processor communication bottlenecks. In this paper, we describe how to enable Parallel Deep Neural Network Training on the IBM Blue Gene/Q (BG/Q) computer system. Specifically, we explore DNN training using the data-parallel Hessian-free 2nd order optimization algorithm. Such an algorithm is particularly well-suited to parallelization across a large set of loosely coupled processors. BG/Q, with its excellent inter-processor communication characteristics, is an ideal match for this type of algorithm. The paper discusses how issues regarding programming model and data-dependent imbalances are addressed. Results on large-scale speech tasks show that the performance on BG/Q scales linearly up to 4,096 processes with no loss in accuracy. This allows us to train neural networks using billions of training examples in a few hours.

Recurrent neural networks (RNNs), including long short-term memory (LSTM) RNNs, have produced state-of-the-art results on a variety of speech recognition tasks. However, these models are often too large in size for deployment on mobile devices with memory and latency constraints. In this work, we study mechanisms for learning compact RNNs and LSTMs via low-rank factorizations and parameter sharing schemes. Our goal is to investigate redundancies in recurrent architectures where compression can be admitted without losing performance. A hybrid strategy of using structured matrices in the bottom layers and shared low-rank factors on the top layers is found to be particularly effective, reducing the parameters of a standard LSTM by 75%, at a small cost of 0.3% increase in WER, on a 2,000-hr English Voice Search task.

Multichannel ASR systems commonly separate speech enhancement, including localization, beamforming and postfiltering, from acoustic modeling. Recently, we explored doing multichannel enhancement jointly with acoustic modeling, where beamforming and frequency decomposition was folded into one layer of the neural network [1, 2]. In this paper, we explore factoring these operations into separate layers in the network. Furthermore, we explore using multi-task learning (MTL) as a proxy for postfiltering, where we train the network to predict "clean" features as well as context-dependent states. We find that with the factored architecture, we can achieve a 10% relative improvement in WER over a single channel and a 5% relative improvement over the unfactored model from [1] on a 2,000-hour Voice Search task. In addition, by incorporating MTL, we can achieve 11% and 7% relative improvements over single channel and unfactored multichannel models, respectively.

Voice Activity Detection (VAD) is an important preprocessing step in any state-of-the-art speech recognition system. Choosing the right set of features and model architecture can be challenging and is an active area of research. In this paper we propose a novel approach to VAD to tackle both feature and model selection jointly. The proposed method is based on a CLDNN (Convolutional, Long Short-Term Memory, Deep Neural Networks) architecture fed directly with the raw waveform. We show that using the raw waveform allows the neural network to learn features directly for the task at hand, which is more powerful than using log-mel features, specially for noisy environments. In addition, using a CLDNN, which takes advantage of both frequency modeling with the CNN and temporal modeling with LSTM, is a much better model for VAD compared to the DNN. The proposed system achieves over 78% relative improvement in False Alarms (FA) at the operating point of 2% False Rejects (FR) on both clean and noisy conditions compared to a DNN of comparable size trained with log-mel features. In addition, we study the impact of the model size and the learned features to provide a better understanding of the proposed architecture.

Multichannel ASR systems commonly use separate modules to perform speech enhancement and acoustic modeling. In this paper, we present an algorithm to do multichannel enhancement jointly with the acoustic model, using a raw waveform convolutional LSTM deep neural network (CLDNN). We will show that our proposed method offers ~5% relative improvement in WER over a log-mel CLDNN trained on multiple channels. Analysis shows that the proposed network learns to be robust to varying angles of arrival for the target speaker, and performs as well as a model that is given oracle knowledge of the true location. Finally, we show that training such a network on inputs captured using multiple (linear) array configurations results in a model that is robust to a range of microphone spacings.

This paper describes a series of experiments to extend the application of Context-Dependent (CD) long short-term memory (LSTM) recurrent neural networks (RNNs) trained with Connectionist Temporal Classification (CTC) and sMBR loss. Our experiments, on a noisy, reverberant voice search task, include training with alternative pronunciations and the application to child speech recognition; combination of multiple models, and convolutional input layers. We also investigate the latency of CTC models and show that constraining forward-backward alignment in training can reduce the delay for a real-time streaming speech recognition system. Finally we investigate transferring knowledge from one network to another through alignments.

We present a novel approach to query-by-example keyword spotting (KWS) using a long short-term memory (LSTM) recurrent neural network-based feature extractor. In our approach, we represent each keyword using a fixed-length feature vector obtained by running the keyword audio through a word-based LSTM acoustic model. We use the activations prior to the softmax layer of the LSTM as our keyword-vector. At runtime, we detect the keyword by extracting the same feature vector from a sliding window and computing a simple similarity score between this test vector and the keyword vector. With clean speech, we achieve 86% relative false rejection rate reduction at 0.5% false alarm rate when compared to a competitive phoneme posteriorgram with dynamic time warping KWS system, while the reduction in the presence of babble noise is 67%. Our system has a small memory footprint, low computational cost, and high precision, making it suitable for on-device applications.

We explore techniques to improve the robustness of small-footprint keyword spotting models based on deep neural networks (DNNs) in the presence of background noise and in far-field conditions. We find that system performance can be improved significantly, with relative improvements up to 75% in far-field conditions, by employing a combination of multi-style training and a proposed novel formulation of automatic gain control (AGC) that estimates the levels of both speech and background noise. Further, we find that these techniques allow us to achieve competitive performance, even when applied to DNNs with an order of magnitude fewer parameters than our base-line.

Both Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) have shown improvements over Deep Neural Networks (DNNs) across a wide variety of speech recognition tasks. CNNs, LSTMs and DNNs are complementary in their modeling capabilities, as CNNs are good at reducing frequency variations, LSTMs are good at temporal modeling, and DNNs are appropriate for mapping features to a more separable space. In this paper, we take advantage of the complementarity of CNNs, LSTMs and DNNs by combining them into one unified architecture. We explore the proposed architecture, which we call CLDNN, on a variety of large vocabulary tasks, varying from 200 to 2,000 hours. We find that the CLDNN provides a 4-6% relative improvement in WER over an LSTM, the strongest of the three individual models.

Recently, Google launched YouTube Kids, a mobile application for children, that uses a speech recognizer built specifically for recognizing children’s speech. In this paper we present techniques we explored to build such a system. We describe the use of a neural network classifier to identify matched acoustic training data, filtering data for language modeling to reduce the chance of producing offensive results. We also compare long short-term memory (LSTM) recurrent networks to convolutional, LSTM, deep neural networks (CLDNN). We found that a CLDNN acoustic model outperforms an LSTM across a variety of different conditions, but does not specifically model child speech relatively better than adult. Overall, these findings allow us to build a successful, state-of-the-art large vocabulary speech recognizer for both children and adults.

We explore using Convolutional Neural Networks (CNNs) for a small-footprint keyword spotting (KWS) task. CNNs are attractive for KWS since they have been shown to outperform DNNs with far fewer parameters. We consider two different applications in our work, one where we limit the number of multiplications of the KWS system, and another where we limit the number of parameters. We present new CNN architectures to address the constraints of each applications. We find that the CNN architectures offer between a 27-44% relative improvement in false reject rate compared to a DNN, while fitting into the constraints of each application.

Learning an acoustic model directly from the raw waveform has been an active area of research. However, waveform-based models have not yet matched the performance of log-mel trained neural networks. We will show that raw wave-form features match the performance of log-mel filterbank ener-gies when used with a state-of-the-art CLDNN acoustic model trained on over 2,000 hours of speech. Specifically, we will show the benefit of the CLDNN, namely the time convolution layer in reducing temporal variations, the frequency convolution layer for preserving locality and reducing frequency variations, as well as the LSTM layers for temporal modeling. In addition, by stacking raw waveform features with log-mel features, we achieve a 3 % relative reduction in word error rate. 1.

We consider the task of building compact deep learning pipelines suitable for deploymenton storage and power constrained mobile devices. We propose a uni-fied framework to learn a broad family of structured parameter matrices that arecharacterized by the notion of low displacement rank. Our structured transformsadmit fast function and gradient evaluation, and span a rich range of parametersharing configurations whose statistical modeling capacity can be explicitly tunedalong a continuum from structured to unstructured. Experimental results showthat these transforms can significantly accelerate inference and forward/backwardpasses during training, and offer superior accuracy-compactness-speed tradeoffsin comparison to a number of existing techniques. In keyword spotting applicationsin mobile speech recognition, our methods are much more effective thanstandard linear low-rank bottleneck layers and nearly retain the performance ofstate of the art models, while providing more than 3.5-fold compression.

Despite their theoretical appeal and grounding in tractable convex optimization techniques, kernel methods are often not the first choice for large-scale speech applications due to their significant memory requirements and computational expense. In recent years, randomized approximate feature maps have emerged as an elegant mechanism to scale-up kernel methods. Still, in practice, a large number of random features is required to obtain acceptable accuracy in predictive tasks. In this paper, we develop two algorithmic schemes to address this computational bottleneck in the context of kernel ridge regression. The first scheme is a specialized distributed block coordinate descent procedure that avoids the explicit materialization of the feature space data matrix, while the second scheme gains efficiency by combining multiple weak random feature models in an ensemble learning framework. We demonstrate that these schemes enable kernel methods to match the performance of state of the art Deep Neural Networks on TIMIT for speech recognition and classification tasks. In particular, we obtain the best classification error rates reported on TIMIT using kernel methods.

State-of-the-art convolutional neural networks (CNNs) typically use a log-mel spectral representation of the speech signal. However, this representation is limited by the spectro-temporal resolution afforded by log-mel filter-banks. A novel technique known as Deep Scattering Spectrum (DSS) addresses this limitation and preserves higher resolution information, while ensuring time warp stability, through the cascaded application of the wavelet-modulus operator. The first order scatter is equivalent to log-mel features and standard CNN modeling techniques can directly be used with these features. However the higher order scatter, which preserves the higher resolution information, presents new challenges in modelling. This paper explores how to effectively use DSS features with CNN acoustic models. Specifically, we identify the effective normalization, neural network topology and regularization techniques to effectively model higher order scatter. The use of these higher order scatter features, in conjunction with CNNs, results in relative improvement of 7% compared to log-mel features on TIMIT, providing a phonetic error rate (PER) of 17.4%, one of the lowest reported PERs to date on this task.

Many features used in speech recognition tasks are hand-crafted and are not always related to the objective at hand, that is minimizing word error rate. Recently, we showed that replacing a perceptually motivated mel-filter bank with a filter bank layer that is learned jointly with the rest of a deep neural network was promising. In this paper, we extend filter learning to a speaker-adapted, state-of-the-art system. First, we incorporate delta learning into the filter learning framework. Second, we incorporate various speaker adaptation techniques, including VTLN warping and speaker identity features. On a 50-hour English Broadcast News task, we show that we can achieve a 5% relative improvement in word error rate (WER) using the filter and delta learning, compared to having a fixed set of filters and deltas. Furthermore, after speaker adaptation, we find that filter and delta learning allows for a 3% relative improvement in WER compared to a state-of-the-art CNN.

We describe a simple modification of neural networks which consists in extending the commonly used linear layer structure to an arbitrary graph structure. This allows us to combine the benefits of convolutional neural networks with the benefits of regular networks. The joint model has only a small increase in parameter size and training and decoding time are virtually unaffected. We report significant improvements over very strong baselines on two LVCSR tasks and one speech activity detection task.

While Deep Neural Networks (DNNs) have achieved tremendous success for LVCSR tasks, training these networks is slow. To date, the most common approach to train DNNs is via stochastic gradient descent (SGD), serially on a single GPU machine. Serial training, coupled with the large number of training parameters and speech data set sizes, makes DNN training very slow for LVCSR tasks. While 2nd order, data-parallel methods have also been explored, these methods are not always faster on CPU clusters due to the large communication cost between processors. In this work, we explore using a specialized hardware/software approach, utilizing a Blue Gene/Q (BG/Q) system, which has thousands of processors and excellent interprocessor communication. We explore using the 2nd order Hessian-free (HF) algorithm for DNN training with BG/Q, for both cross-entropy and sequence training of DNNs. Results on three LVCSR tasks indicate that using HF with BG/Q offers up to an 11x speedup, as well as an improved word error rate (WER), compared to SGD on a GPU.

Log-mel filterbank features, which are commonly used features for CNNs, can remove higher-resolution information from the speech signal. A novel technique, known as Deep Scattering Spectrum (DSS), addresses this issue and looks to preserve this information. DSS features have shown promise on TIMIT, both for classification and recognition. In this paper, we extend the use of DSS features for LVCSR tasks. First, we explore the optimal multi-resolution time and frequency scattering operations for LVCSR tasks. Next, we explore techniques to reduce the dimension of the DSS features. We also incorporate speaker adaptation techniques into the DSS features. Results on a 50 and 430 hour English Broadcast News task show that the DSS features provide between a 4-7% relative improvement in WER over log-mel features, within a state-of-the-art CNN framework which incorporates speaker-adaptation and sequence training. Finally, we show that DSS features are similar to multi-resolution log-mel + MFCCs, and similar improvements can be obtained with this representation.

Deep Neural Networks (DNNs) have recently been shown to significantly outperform existing machine learning techniques in several pattern recognition tasks. DNNs are the state-of-the-art models used in image recognition, object detection, classification and tracking, and speech and language processing applications. The biggest drawback to DNNs has been the enormous cost in computation and time taken to train the parameters of the networks - often a tenfold increase relative to conventional technologies. Such training time costs can be mitigated by the application of parallel computing algorithms and architectures. However, these algorithms often run into difficulties because of the cost of inter-processor communication bottlenecks. In this paper, we describe how to enable Parallel Deep Neural Network Training on the IBM Blue Gene/Q (BG/Q) computer system. Specifically, we explore DNN training using the data parallel Hessian-free 2nd order optimization algorithm. Such an algorithm is particularly well-suited to parallelization across a large set of loosely coupled processors. BG/Q, with its excellent inter-processor communication characteristics, is an ideal match for this type of algorithm. The paper discusses how issues regarding programming model and data-dependent imbalances are addressed. Results on large-scale speech tasks show that the performance on BG/Q scales linearly up to 4096 processes with no loss in accuracy. This allows us to train neural networks using billions of training examples in a few hours.

Hessian-free training has become a popular parallel second order 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 1.5× speed-up, whereas, on a 300-hr Switchboard task, these techniques provide over a 2.3× speedup, with no loss in WER. These results suggest that even further speed-up is expected, as problems scale and complexity grows.

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.

Mel-filter banks are commonly used in speech recognition, as they are motivated from theory related to speech production and perception. While features derived from mel-filter banks are quite popular, we argue that this filter bank is not really an appropriate choice as it is not learned for the objective at hand, i.e. speech recognition. In this paper, we explore replacing the filter bank with a filter bank layer that is learned jointly with the rest of a deep neural network. Thus, the filter bank is learned to minimize cross-entropy, which is more closely tied to the speech recognition objective. On a 50-hour English Broadcast News task, we show that we can achieve a 5% relative improvement in word error rate (WER) using the filter bank learning approach, compared to having a fixed set of filters.

Automatic speech recognition is a core component of many applications, including keyword search. In this paper we describe experiments on acoustic modeling, language modeling, and decoding for keyword search on a Cantonese conversational telephony corpus collected as part of the IARPA Babel program. We show that acoustic modeling techniques such as the bootstrapped-and-restructured model and deep neural network acoustic model significantly outperform a state-of-the-art baseline GMM/HMM model, in terms of both recognition performance and keyword search performance, with improvements of up to 11% relative character error rate reduction and 31% relative maximum term weighted value improvement. We show that while an interpolated Model M and neural network LM improve recognition performance, they do not improve keyword search results; however, the advanced LM does reduce the size of the keyword search index. Finally, we show that a simple form of automatically adapted keyword search performs 16% better than a preindexed search system, indicating that out-of-vocabulary search is still a challenge.

Recently, pre-trained deep neural networks (DNNs) have outperformed traditional acoustic models based on Gaussian mixture models (GMMs) on a variety of large vocabulary speech recognition benchmarks. Deep neural nets have also achieved excellent results on various computer vision tasks using a random “dropout” procedure that drastically improves generalization error by randomly omitting a fraction of the hidden units in all layers. Since dropout helps avoid over-fitting, it has also been successful on a small-scale phone recognition task using larger neural nets. However, training deep neural net acoustic models for large vocabulary speech recognition takes a very long time and dropout is likely to only increase training time. Neural networks with rectified linear unit (ReLU) non-linearities have been highly successful for computer vision tasks and proved faster to train than standard sigmoid units, sometimes also improving discriminative performance. In this work, we show on a 50-hour English Broadcast News task that modified deep neural networks using ReLUs trained with dropout during frame level training provide an 4.2% relative improvement over a DNN trained with sigmoid units, and a 14.4% relative improvement over a strong GMM/HMM system. We were able to obtain our results with minimal human hyper-parameter tuning using publicly available Bayesian optimization code.

Several methods have been proposed recently for modeling posterior representations derived from local classifiers [1, 2]. In recent work, Sainath et al. have proposed the use of a tied-mixture-based posterior modeling approach [3] to enhance exemplar-based posterior representations for phone recognition tasks. In this work, we conduct a detailed evaluation to determine the effectiveness of this technique on three representative posterior systems. In addition, we propose and evaluate an alternative discriminative formulation of the posterior modeling objective function that seeks to minimize framelevel errors. In experimental evaluations on the TIMIT corpus, we find that posterior modeling results in relative phone error rate (PER) reductions of between 1.1-5.5% across the systems tested. In fact, using Spif-NN [4, 3] posteriors, we are able to achieve a PER of 18.5; to the best of our knowledge, this is the best result reported in the literature to date. minimize framelevel errors.

While Deep Neural Networks (DNNs) have achieved tremendous success for large vocabulary continuous speech recognition (LVCSR) tasks, training of these networks is slow. One reason is that DNNs are trained with a large number of training parameters (i.e., 10–50 million). Because networks are trained with a large number of output targets to achieve good performance, the majority of these parameters are in the final weight layer. In this paper, we propose a low-rank matrix factorization of the final weight layer. We apply this low-rank technique to DNNs for both acoustic modeling and language modeling. We show on three different LVCSR tasks ranging between 50–400 hrs, that a low-rank factorization reduces the number of parameters of the network by 30–50%. This results in roughly an equivalent reduction in training time, without a significant loss in final recognition accuracy, compared to a full-rank representation.

Convolutional Neural Networks (CNNs) are an alternative type of neural network that can be used to reduce spectral variations and model spectral correlations which exist in signals. Since speech signals exhibit both of these properties, CNNs are a more effective model for speech compared to Deep Neural Networks (DNNs). In this paper, we explore applying CNNs to large vocabulary speech tasks. First, we determine the appropriate architecture to make CNNs effective compared to DNNs for LVCSR tasks. Specifically, we focus on how many convolutional layers are needed, what is the optimal number of hidden units, what is the best pooling strategy, and the best input feature type for CNNs. We then explore the behavior of neural network features extracted from CNNs on a variety of LVCSR tasks, comparing CNNs to DNNs and GMMs. We find that CNNs offer between a 13-30% relative improvement over GMMs, and a 4-12% relative improvement over DNNs, on a 400-hr Broadcast News and 300-hr Switchboard task.

While Deep Neural Networks (DNNs) have achieved tremendous success for large vocabulary continuous speech recognition (LVCSR) tasks, training these networks is slow. Even to date, the most common approach to train DNNs is via stochastic gradient descent, serially on one machine. Serial training, coupled with the large number of training parameters (i.e., 10-50 million) and speech data set sizes (i.e., 20-100 million training points) makes DNN training very slow for LVCSR tasks. In this work, we explore a variety of different optimization techniques to improve DNN training speed. This includes parallelization of the gradient computation during cross-entropy and sequence training, as well as reducing the number of parameters in the network using a low-rank matrix factorization. Applying the proposed optimization techniques, we show that DNN training can be sped up by a factor of 3 on a 50-hour English Broadcast News (BN) task with no loss in accuracy. Furthermore, using the proposed techniques, we are able to train DNNs on a 300-hr Switchboard (SWB) task and a 400-hr English BN task, showing improvements between 9-30% relative over a state-of-the art GMM/HMM system while the number of parameters of the DNN is smaller than the GMM/HMM system.

This paper presents a strategy for efficiently selecting informative data from large corpora of untranscribed speech. Confidence-based selection methods (i.e., selecting utterances we are least confident about) have been a popular approach, though they only look at the top hypothesis when selecting utterances and tend to select outliers, therefore, not always improving overall recognition accuracy. Alternatively, we propose a method for selecting data looking at competing hypothesis by computing entropy of N-best hypothesis decoded by the baseline acoustic model. In addition we address the issue of outliers by calculating how representative a specific utterance is to all other unselected utterances via a tf-idf score. Experiments show that N-best entropy based selection (%relative 5.8 in 400-hour corpus) outperformed other conventional selection strategies; confidence based and lattice entropy based, and that tf-idf based representativeness improved the model further (%relative 6.2). A comparison with random selection is also presented. Finally model size impact is discussed.

Restricted Boltzmann Machines (RBM) continue to be a popular methodology to pre-train weights of Deep Belief Networks (DBNs). However, the RBM objective function cannot be maximized directly. Therefore, it is not clear what function to monitor when deciding to stop the training, leading to a challenge in managing the computational costs. The Sparse Encoding Symmetric Machine (SESM) has been suggested as an alternative method for pre-training. By placing a sparseness term on the NN output codebook, SESM allows the objective function to be optimized directly and reliably be monitored as an indicator to stop the training. In this paper, we explore SESM to pre-train DBNs and apply this the first time to speech recognition. First, we provide a detailed analysis comparing the behavior of SESM and RBM. Second, we compare the performance of SESM pre-trained and RBM pre-trained DBNs on TIMIT and a 50 hour English Broadcast News task. Results indicate that pre-trained DBNs using SESM and RBMs achieve comparable performance and outperform randomly initialized DBNs with SESM providing a much easier stopping criterion relative to RBM.

Neural network (NN) bottleneck (BN) features are typically created by training a NN with a middle bottleneck layer. Recently, an alternative structure was proposed which trains a NN with a constant number of hidden units to predict output targets, and then reduces the dimensionality of these output probabilities through an auto-encoder, to create auto-encoder bottleneck (AE-BN) features. The benefit of placing the BN after the posterior estimation network is that it avoids the loss in frame classification accuracy incurred by networks that place the BN before the softmax. In this work, we investigate the use of pre-training when creating AE-BN features. Our experiments indicate that with the AE-BN architecture, pre-trained and deeper NNs produce better AE-BN features. On a 50-hour English Broadcast News task, the AE-BN features provide over a 1% absolute improvement compared to a state-of-the-art GMM/HMM with a WER of 18.8% and pre-trained NN hybrid system with a WER of 18.4%. In addition, on a larger 430-hour Broadcast News task, AE-BN features provide a 0.5% absolute improvement over a strong GMM/HMM baseline with a WER of 16.0%. Finally, system combination with the GMM/HMM baseline and AE-BN systems provides an additional 0.5% absolute on 430 hours over the AE-BN system alone, yielding a final WER of 15.0%.

To date, there has been limited work in applying Deep Belief Networks (DBNs) for acoustic modeling in LVCSR tasks, with past work using standard speech features. However, a typical LVCSR system makes use of both feature and model-space speaker adaptation and discriminative training. This paper explores the performance of DBNs in a state-of-the-art LVCSR system, showing improvements over Multi-Layer Perceptrons (MLPs) and GMM/HMMs across a variety of features on an English Broadcast News task. In addition, we provide a recipe for data parallelization of DBN training, showing that data parallelization can provide linear speed-up in the number of machines, without impacting WER.

In this paper, we propose a novel exemplar based technique for classification problems where for every new test sample the classification model is re-estimated from a subset of relevant samples of the training data.We formulate the exemplar-based classification paradigm as a sparse representation (SR) problem, and explore the use of convex hull constraints to enforce both regularization and sparsity. Finally, we utilize the Extended Baum-Welch (EBW) optimization technique to solve the SR problem. We explore our proposed methodology on the TIMIT phonetic classification task, showing that our proposed method offers statistically significant improvements over common classification methods, and provides an accuracy of 82.9%, the best single-classifier number reported to date.

We introduce the Line Search A-Function (LSAF) technique that generalizes the Extended-Baum Welch technique in order to provide an effective optimization technique for a broader set of functions. We show how LSAF can be applied to functions of various probability density and distribution functions by demonstrating that these probability functions have an A-function. We also show that sparse representation problems (SR) that use 11 or combination of 11/12 regularization norms can also be efficiently optimized through an A-function derived for their objective functions. We will demonstrate the efficiency of LSAF for SR problems through simulations by comparing it with Approximate Bayesian Compressive Sensing method that we recently applied to speech recognition.

Deep Belief Networks (DBNs) are multi-layer generative models. They can be trained to model windows of coefficients extracted from speech and they discover multiple layers of features that capture the higher-order statistical structure of the data. These features can be used to initialize the hidden units of a feed-forward neural network that is then trained to predict the HMM state for the central frame of the window. Initializing with features that are good at generating speech makes the neural network perform much better than initializing with random weights. DBNs have already been used successfully for phone recognition with input coefficients that are MFCCs or filterbank outputs. In this paper, we demonstrate that they work even better when their inputs are speaker adaptive, discriminative features. On the standard TIMIT corpus, they give phone error rates of 19.6% using monophone HMMs and a bigram language model and 19.4% using monophone HMMs and a trigram language model.

Exemplar-based techniques, such as k-nearest neighbors (kNNs) and Sparse Representations (SRs), can be used to model a test sample from a few training points in a dictionary set. In past work, we have shown that using a SR approach for phonetic classification allows for a higher accuracy than other classification techniques. These phones are the basic units of speech to be recognized. Motivated by this result, we create a new dictionary which is a function of the phonetic labels of the original dictionary. The SR method now selects relevant samples from this new dictionary to create a new feature representation of the test sample, where the new feature is better linked to the actual units to be recognized. We will refer to these new features as Spif. We present results using these new Spiffeatures in a Hidden Markov Model (HMM) framework for speech recognition. We find that the Spiffeatures allow for a 2.9% relative reduction in Phonetic Error Rate (PER) on the TIMIT phonetic recognition task. Furthermore, we find that the Spiffeatures allow for a 4.8% relative improvement in Word Error Rate (WER) on a large vocabulary 50 hour Broadcast News task.

Gradient boosting is a flexible machine learning technique that produces accurate predictions by combining many weak learners. In this work, we investigate its use in two applications, where we show the advantage of loss functions that are designed specifically for optimizing application objectives. We also extend the original gradient boosting algorithm with Newton-Raphson method to speed up learning. In the experiments, we demonstrate that the use of gradient boosting and application specific loss functions results in a relative improvement of 0.8% over an 82.6% baseline on the CoNLL 2003 named entity recognition task. We also show that this novel framework is useful in identifying regions of high word error rate (WER) and can provide up to 20% relative improvement depending on the chosen operating point.

The use of exemplar-based methods, such as support vector machines (SVMs), k-nearest neighbors (kNNs) and sparse representations (SRs), in speech recognition has thus far been limited. Exemplar-based techniques utilize information about individual training examples and are computationally expensive, making it particularly difficult to investigate these methods on large-vocabulary continuous speech recognition (LVCSR) tasks. While research in LVCSR provides a good testbed to tackle real-world speech recognition problems, research in this area suffers from two main drawbacks. First, the overall complexity of an LVCSR system makes error analysis quite difficult. Second, exploring new research ideas on LVCSR tasks involves training and testing state-of-the-art LVCSR systems, which can render a large turnaround time. This makes a small vocabulary task such as TIMIT more appealing. TIMIT provides a phonetically rich and hand-labeled corpus that allows easy insight into new algorithms. However, research ideas explored for small vocabulary tasks do not always provide gains on LVCSR systems. In this paper, we combine the advantages of using both small and large vocabulary tasks by taking well-established techniques used in LVCSR systems and applying them on TIMIT to establish a new baseline. We then utilize these existing LVCSR techniques in creating a novel set of exemplar-based sparse representation (SR) features. Using these existing LVCSR techniques, we achieve a phonetic error rate (PER) of 19.4% on the TIMIT task. The additional use of SR features reduce the PER to 18.6%. We then explore applying the SR features to a large vocabulary Broadcast News task, where we achieve a 0.3% absolute reduction in word error rate (WER).

Compressive sensing (CS) is a popular technique used to reconstruct a signal from few training examples, a problem which arises in many machine learning applications. In this paper, we introduce a technique to guarantee that our data obeys certain isometric properties. In addition, we introduce a bayesian approach to compressive sensing, which we call ABCS, allowing us to obtain complete statistics for estimated parameters. We apply these ideas to fMRI classification and find that by isometrically transforming our data, significant improvements in classification accuracy can be achieved using the LASSO and Dantzig selector methods, two standard techniques used in CS. In addition, applying the ABCS method offers improvements in classification accuracy over both LASSO and Dantzig. Finally, we find that applying both the ABCS method together with isometric transformations, we are able to achieve an error rate of 0.0%.

In this paper, we introduce a novel Bayesian compressive sensing (CS) technique for phonetic classification. CS is often used to characterize a signal from a few support training examples, similar to k-nearest neighbor (kNN) and Support Vector Machines (SVMs). However, unlike SVMs and kNNs, CS allows the number of supports to be adapted to the specific signal being characterized. On the TIMIT phonetic classification task, we find that our CS method outperforms the SVM, kNN and Gaussian Mixture Model (GMM) methods. Our CS method achieves an accuracy of 80.01%, one of the best reported result in the literature to date.

One long-standing challenge in robotics is the realization of mobile autonomous robots able to operate safely in existing human workplaces in a way that their presence is accepted by the human occupants. We describe the development of a multi-ton robotic forklift intended to operate alongside human personnel, handling palletized materials within existing, busy, semi-structured outdoor storage facilities. The system has three principal novel characteristics. The first is a multimodal tablet that enables human supervisors to use speech and pen-based gestures to assign tasks to the forklift, including manipulation, transport, and placement of palletized cargo. Second, the robot operates in minimally-prepared, semi-structured environments, in which the forklift handles variable palletized cargo using only local sensing (and no reliance on GPS), and transports it while interacting with other moving vehicles. Third, the robot operates in close proximity to people, including its human supervisor, other pedestrians who may cross or block its path, and forklift operators who may climb inside the robot and operate it manually. This is made possible by novel interaction mechanisms that facilitate safe, effective operation around people. We describe the architecture and implementation of the system, indicating how real-world operational requirements motivated the development of the key subsystems, and provide qualitative and quantitative descriptions of the robot operating in real settings.

Sparse representation phone identification features (SPIF) is a recently developed technique to obtain an estimate of phone posterior probabilities conditioned on an acoustic feature vector. In this paper, we explore incorporating SPIF phone posterior probability estimates in large vocabulary continuous speech recognition (LVCSR) task by including them as additional features of exponential densities that model the HMM state emission likelihoods. We compare our proposed approach to a number of other well known methods of combining feature streams or multiple LVCSR systems. Our experiments show that using exponential models to combine features results in a word error rate reduction of 0.5% absolute (18.7% down to 18.2%); this is comparable to best error rate reduction obtained from system combination methods, but without having to build multiple systems or tune the system combination weights.

The use of exemplar-based techniques for both speech classification and recognition tasks has become increasingly popular in recent years. However, the notion of why sparseness is important for exemplar-based speech processing has been relatively unexplored. In addition, little analysis has been done in speech processing on the appropriateness of different types of sparsity regularization constraints. The goal of this paper is to answer the above two questions, both through mathematically analyzing different sparseness methods and also comparing these approaches for phonetic classification in TIMIT.

Sparse representations (SRs) are often used to characterize a test signal using few support training examples, and allow the number of supports to be adapted to the specific signal being categorized. Given the good performance of SRs compared to other classifiers for both image classification and phonetic clas-sification, in this paper, we extended the use of SRs for text classification, a method which has thus far not been explored for this domain. Specifically, we demonstrate how sparse repre-sentations can be used for text classification and how their per-formance varies with the vocabulary size of the documents. In addition, we also show that this method offers promising results over the Naive Bayes (NB) classifier, a standard baseline classi-fier used for text categorization, thus introducing an alternative class of methods for text categorization. 1.

In this paper, we explore the use of exemplar-based sparse representations (SRs) to map test features into the linear span of training examples. We show that the frame classification ac-curacy with these new features is 1.3 % higher than a Gaussian Mixture Model (GMM), showing that not only do SRs move test features closer to training, but also move the features closer to the correct class. Given these new SR features, we train up a Hidden Markov Model (HMM) on these features and per-form recognition. On the TIMIT corpus, we show that apply-ing the SR features on top of our best discriminatively trained system allows for a 0.7 % absolute reduction in phonetic er-ror rate (PER), from 19.9 % to 19.2%. In fact, after applying model adaptation we reduce the PER to 19.0%, the best re-sults on TIMIT to date. Furthermore, on a large vocabulary 50 hour broadcast news task, we achieve a reduction in word error rate (WER) of 0.3 % absolute, demonstrating the benefit of this method for large vocabulary speech recognition. 1.

Most speech recognizers do not differentiate between reliable and unreliable portions of the speech signal during search. As a result, most of the search effort is concentrated in unreliable areas. Island-driven search addresses this problem by first identifying reliable islands and directing the search out from these islands towards unreliable gaps. In this paper, we develop a technique to detect islands from knowledge of hypothesized broad phonetic classes (BPCs). Using this island/gap knowledge, we explore a method to prune the search space to limit computational effort in unreliable areas. In addition, we also investigate scoring less detailed BPC models in gap regions and more detailed phonetic models in islands. Experiments on both small and large scale vocabulary tasks indicate that our island-driven search strategy results in an improvement in recognition accuracy and computation time.

While research in large vocabulary continuous speech recognition (LVCSR) has sparked the development of many state of the art research ideas, research in this domain suffers from two main drawbacks. First, because of the large number of parameters and poorly labeled transcriptions, gaining insight into further improvements based on error analysis is very difficult. Second, LVCSR systems often take a significantly longer time to train and test new research ideas compared to small vocabulary tasks. A small vocabulary task like TIMIT provides a phonetically rich and hand-labeled corpus and offers a good test bed to study algorithmic improvements. However, oftentimes research ideas explored for small vocabulary tasks do not always provide gains on LVCSR systems. In this paper, we address these issues by taking the standard "recipe" used in typical LVCSR systems and applying it to the TIMIT phonetic recognition corpus, which provides a standard benchmark to compare methods. We find that at the speaker-independent (SI) level, our results offer comparable performance to other SI HMM systems. By taking advantage of speaker adaptation and discriminative training techniques commonly used in LVCSR systems, we achieve an error rate of 20%, the best results reported on the TIMIT task to date, moving us closer to the human reported phonetic recognition error rate of 15%. We propose the use of this system as the baseline for future research and believe that it will serve as a good framework to explore ideas that will carry over to LVCSR systems.

The extended Baum-Welch (EBW) transformations is one of a variety of techniques to estimate parameters of Gaussian mixture models. In this paper, we provide a theoretical framework for general parameter estimation and show the relationship between these different techniques. We introduce a general family of model parameter updates that generalizes a Baum-Welch (BW) recursive process to an arbitrary objective function of Gaussian Mixture Models, and show how other common parameter estimation techniques belong to this family of model update rules. Furthermore, we formulate the construction of an even more general family of update rules that has any specified value as a gradient steepness which belongs to the family of EBW gradient steepness, measuring how much an initial model is moved to an estimated updated model.

In many pattern recognition tasks, given some input data and a family of models, the "best" model is defined as the one which maximizes the likelihood of the data given the model. Extended Baum- Welch (EBW) transformations are most commonly used as a discriminative technique for estimating parameters of Gaussian mixtures. In this paper, we use the EBW transformations to derive a novel gradient steepness measurement to find which model best explains the data. We use this gradient measurement to derive a variety of EBW metrics to explain model fit to the data. We apply these EBW metrics to audio segmentation via Hidden Markov Models (HMMs) and show that our gradient steepness measurement is robust across different EBW metrics and model complexities.

We demonstrate the generalizability of the Extended Baum-Welch (EBW) algorithm not only for HMM parameter estimation but for decoding as well. We show that there can exist a general function associated with the objective function under EBW that reduces to the well-known auxiliary function used in the Baum-Welch algorithm for maximum likelihood estimates. We generalize representation for the updates of model parameters by making use of a differentiable function (such as arithmetic or geometric mean) on the updated and current model parameters and describe their effect on the learning rate during HMM parameter estimation. Improvements on speech recognition tasks are also presented here. 1.

In this paper, we compare speech recognition performance using broad phoneticallyand acoustically-motivated units as a pre-processor in designing a novel noise robust landmark detection and segmentation algorithm. We introduce a cluster evaluation method to measure acoustic unit cluster quality. On the noisy TIMIT task, we find that the acoustic and phonetic segmentation approaches offer significant improvements over two baseline methods used in the SUMMIT segment-based speech recognizer, a sinusoidal model method and a spectral change approach. In addition, we find that the acoustic method has much faster computation time in stationary noises, while the phonetic approach is faster in non-stationary noise conditions.

In many pattern recognition tasks, given some input data and a model, a probabilistic likelihood score is often computed to measure how well the model describes the data. Extended Baum-Welch (EBW) transformations are most commonly used as a discriminative technique for estimating parameters of Gaussian mixtures, though recently they have been used to derive a gradient steepness measurement to evaluate the quality of the model to match the distribution of the data. In this paper, we explore applying the EBW gradient steepness metric in the context of Hidden Markov Models (HMMs) for recognition of broad phonetic classes and present a detailed analysis and results on the use of this gradient metric on the TIMIT corpus. We find that our gradient metric is able to outperform the baseline likelihood method, and offers improvements in noisy conditions.

Audio segmentation has applications in a variety of contexts, such as audio information retrieval, automatic sound analysis, and as a pre-processing step in speech recognition. Extended Baum-Welch (EBW) transformations are most commonly used as a discriminative technique for estimating parameters of Gaussian mixtures. In this paper, we derive an unsupervised audio segmentation approach using these transformations. We find that our algorithm outperforms both the Bayesian information criterion (BIC) and cumulative sum (CUSUM) segmentation methods. In particular, our EBW segmentation algorithm provides improvements over the baseline approaches in detecting landmarks of short duration and minimizing landmark oversegmentation. In addition, we show that the EBW approach provides faster computation compared to the baseline methods.

Audio classification has applications in a variety of contexts, such as automatic sound analysis, supervised audio segmentation and in audio information search and retrieval. Extended Baum-Welch (EBW) transformations are most commonly used as a discriminative technique for estimating parameters of Gaussian mixtures, though recently they have been applied in unsupervised audio segmentation. In this paper, we extend the use of these transformations to derive an audio classification algorithm. We find that our method outperforms both the Support Vector Machine (SVM) and Gaussian Mixture Model (GMM) likelihood classification methods. Index Terms: audio classification, gradient methods 1.

In this paper, we present a noise robust landmark detection and segmentation algorithm using a sinusoidal model representation of speech. We compare the performance of our approach under noisy conditions against two segmentation methods used in the SUMMIT segment-based speech recognizer, a full segmentation approach and an approach that detects segment boundaries based on spectral change. The word error rate of the spectral change segmentation method degrades rapidly in the presence of noise, while the sinusoidal and full segmentation models degrade more gracefully. However, the full segmentation method requires the largest computation time of the three approaches. We find that our new algorithm provides the best tradeoff between word accuracy and computation time of the three methods. Furthermore, we find that our model is robust when speech is contaminated by various noise types