Yangyang Shi

Haichuan Yang, Zhaojun Yang, Li Wan et al. · amazon-science

This paper proposes a hardware-efficient architecture, Linearized Convolution Network (LiCo-Net) for keyword spotting. It is optimized specifically for low-power processor units like microcontrollers. ML operators exhibit heterogeneous efficiency profiles on power-efficient hardware. Given the exact theoretical computation cost, int8 operators are more computation-effective than float operators, and linear layers are often more efficient than other layers. The proposed LiCo-Net is a dual-phase system that uses the efficient int8 linear operators at the inference phase and applies streaming convolutions at the training phase to maintain a high model capacity. The experimental results show that LiCo-Net outperforms single-value decomposition filter (SVDF) on hardware efficiency with on-par detection performance. Compared to SVDF, LiCo-Net reduces cycles by 40% on HiFi4 DSP.

Florian L. Kreyssig, Yangyang Shi, Jinxi Guo et al.

Self-supervised learning via masked prediction pre-training (MPPT) has shown impressive performance on a range of speech-processing tasks. This paper proposes a method to bias self-supervised learning towards a specific task. The core idea is to slightly finetune the model that is used to obtain the target sequence. This leads to better performance and a substantial increase in training speed. Furthermore, this paper proposes a variant of MPPT that allows low-footprint streaming models to be trained effectively by computing the MPPT loss on masked and unmasked frames. These approaches are evaluated for automatic speech recognition on the Librispeech corpus, where 100 hours of data served as the labelled data and 860 hours as the unlabelled data. The biased training outperforms the unbiased training by 15.5% after 250k updates and 23.8% after 100k updates on test-other. For the streaming models, the pre-training approach yields a reduction in word error rate of 44.1%.

Mei-Yuh Hwang, Yangyang Shi, Ankit Ramchandani et al.

This paper discusses the challenges of optical character recognition (OCR) on natural scenes, which is harder than OCR on documents due to the wild content and various image backgrounds. We propose to uniformly use word error rates (WER) as a new measurement for evaluating scene-text OCR, both end-to-end (e2e) performance and individual system component performances. Particularly for the e2e metric, we name it DISGO WER as it considers Deletion, Insertion, Substitution, and Grouping/Ordering errors. Finally we propose to utilize the concept of super blocks to automatically compute BLEU scores for e2e OCR machine translation. The small SCUT public test set is used to demonstrate WER performance by a modularized OCR system.

Yang Li, Liangzhen Lai, Yuan Shangguan et al.

Transformer-based models excel in speech recognition. Existing efforts to optimize Transformer inference, typically for long-context applications, center on simplifying attention score calculations. However, streaming speech recognition models usually process a limited number of tokens each time, making attention score calculation less of a bottleneck. Instead, the bottleneck lies in the linear projection layers of multi-head attention and feedforward networks, constituting a substantial portion of the model size and contributing significantly to computation, memory, and power usage. To address this bottleneck, we propose folding attention, a technique targeting these linear layers, significantly reducing model size and improving memory and power efficiency. Experiments on on-device Transformer-based streaming speech recognition models show that folding attention reduces model size (and corresponding memory consumption) by up to 24% and power consumption by up to 23%, all without compromising model accuracy or computation overhead.

Yangyang Shi, Gael Le Lan, Varun Nagaraja et al.

This paper presents an innovative approach to enhance control over audio generation by emphasizing the alignment between audio and text representations during model training. In the context of language model-based audio generation, the model leverages input from both textual and audio token representations to predict subsequent audio tokens. However, the current configuration lacks explicit regularization to ensure the alignment between the chosen text representation and the language model's predictions. Our proposal involves the incorporation of audio and text representation regularization, particularly during the classifier-free guidance (CFG) phase, where the text condition is excluded from cross attention during language model training. The aim of this proposed representation regularization is to minimize discrepancies in audio and text similarity compared to other samples within the same training batch. Experimental results on both music and audio generation tasks demonstrate that our proposed methods lead to improvements in objective metrics for both audio and music generation, as well as an enhancement in the human perception for audio generation.

Ernie Chang, Muhammad Hassan Rashid, Pin-Jie Lin et al.

Knowing exactly how many data points need to be labeled to achieve a certain model performance is a hugely beneficial step towards reducing the overall budgets for annotation. It pertains to both active learning and traditional data annotation, and is particularly beneficial for low resource scenarios. Nevertheless, it remains a largely under-explored area of research in NLP. We therefore explored various techniques for estimating the training sample size necessary to achieve a targeted performance value. We derived a simple yet effective approach to predict the maximum achievable model performance based on small amount of training samples - which serves as an early indicator during data annotation for data quality and sample size determination. We performed ablation studies on four language understanding tasks, and showed that the proposed approach allows us to forecast model performance within a small margin of mean absolute error (~ 0.9%) with only 10% data.

Chunxi Liu, Yuan Shangguan, Haichuan Yang et al.

There is growing interest in unifying the streaming and full-context automatic speech recognition (ASR) networks into a single end-to-end ASR model to simplify the model training and deployment for both use cases. While in real-world ASR applications, the streaming ASR models typically operate under more storage and computational constraints - e.g., on embedded devices - than any server-side full-context models. Motivated by the recent progress in Omni-sparsity supernet training, where multiple subnetworks are jointly optimized in one single model, this work aims to jointly learn a compact sparse on-device streaming ASR model, and a large dense server non-streaming model, in a single supernet. Next, we present that, performing supernet training on both wav2vec 2.0 self-supervised learning and supervised ASR fine-tuning can not only substantially improve the large non-streaming model as shown in prior works, and also be able to improve the compact sparse streaming model.

Ernie Chang, Sidd Srinivasan, Mahi Luthra et al.

Text-to-audio generation (TTA) produces audio from a text description, learning from pairs of audio samples and hand-annotated text. However, commercializing audio generation is challenging as user-input prompts are often under-specified when compared to text descriptions used to train TTA models. In this work, we treat TTA models as a ``blackbox'' and address the user prompt challenge with two key insights: (1) User prompts are generally under-specified, leading to a large alignment gap between user prompts and training prompts. (2) There is a distribution of audio descriptions for which TTA models are better at generating higher quality audio, which we refer to as ``audionese''. To this end, we rewrite prompts with instruction-tuned models and propose utilizing text-audio alignment as feedback signals via margin ranking learning for audio improvements. On both objective and subjective human evaluations, we observed marked improvements in both text-audio alignment and music audio quality.

Ernie Chang, Pin-Jie Lin, Yang Li et al.

Distributional shift is a central challenge in the deployment of machine learning models as they can be ill-equipped for real-world data. This is particularly evident in text-to-audio generation where the encoded representations are easily undermined by unseen prompts, which leads to the degradation of generated audio -- the limited set of the text-audio pairs remains inadequate for conditional audio generation in the wild as user prompts are under-specified. In particular, we observe a consistent audio quality degradation in generated audio samples with user prompts, as opposed to training set prompts. To this end, we present a retrieval-based in-context prompt editing framework that leverages the training captions as demonstrative exemplars to revisit the user prompts. We show that the framework enhanced the audio quality across the set of collected user prompts, which were edited with reference to the training captions as exemplars.

Zechun Liu, Changsheng Zhao, Forrest Iandola et al.

This paper addresses the growing need for efficient large language models (LLMs) on mobile devices, driven by increasing cloud costs and latency concerns. We focus on designing top-quality LLMs with fewer than a billion parameters, a practical choice for mobile deployment. Contrary to prevailing belief emphasizing the pivotal role of data and parameter quantity in determining model quality, our investigation underscores the significance of model architecture for sub-billion scale LLMs. Leveraging deep and thin architectures, coupled with embedding sharing and grouped-query attention mechanisms, we establish a strong baseline network denoted as MobileLLM, which attains a remarkable 2.7%/4.3% accuracy boost over preceding 125M/350M state-of-the-art models. Additionally, we propose an immediate block-wise weight-sharing approach with no increase in model size and only marginal latency overhead. The resultant models, denoted as MobileLLM-LS, demonstrate a further accuracy enhancement of 0.7%/0.8% than MobileLLM 125M/350M. Moreover, MobileLLM model family shows significant improvements compared to previous sub-billion models on chat benchmarks, and demonstrates close correctness to LLaMA-v2 7B in API calling tasks, highlighting the capability of small models for common on-device use cases.

Zhipeng Cai, Ching-Feng Yeh, Hu Xu et al. · meta-ai, mit

Vision language models (VLMs) can flexibly address various vision tasks through text interactions. Although successful in semantic understanding, state-of-the-art VLMs including GPT-5 still struggle in understanding 3D from 2D inputs. On the other hand, expert pure vision models achieve super-human accuracy in metric depth estimation, a key 3D understanding task. However, they require task-specific architectures and losses. Such difference motivates us to ask: Can VLMs reach expert-level accuracy without architecture or loss change? We take per-pixel metric depth estimation as the representative task and show that the answer is yes! Surprisingly, comprehensive analysis shows that text-based supervised-finetuning with sparse labels is sufficient for VLMs to unlock strong 3D understanding, no dense prediction head or complex regression/regularization loss is needed. The bottleneck for VLMs lies actually in pixel reference and cross-dataset camera ambiguity, which we address through visual prompting and intrinsic-conditioned augmentation. With much smaller models, our method DepthLM surpasses the accuracy of most advanced VLMs by over 2x, making VLMs for the first time comparable with pure vision models. Interestingly, without explicit enforcement during training, VLMs trained with DepthLM naturally avoids over-smoothing, having much fewer flying points at boundary regions than pure vision models. The simplicity of DepthLM also enables a single VLM to cover various 3D tasks beyond metric depth. Our code and model will be released at the link below.

Xiaoyan Kui, Canwei Liu, Qinsong Li et al.

Kolmogorov-Arnold Networks (KANs) are highly effective in long-term time series forecasting due to their ability to efficiently represent nonlinear relationships and exhibit local plasticity. However, prior research on KANs has predominantly focused on the time domain, neglecting the potential of the frequency domain. The frequency domain of time series data reveals recurring patterns and periodic behaviors, which complement the temporal information captured in the time domain. To address this gap, we explore the application of KANs in the frequency domain for long-term time series forecasting. By leveraging KANs' adaptive activation functions and their comprehensive representation of signals in the frequency domain, we can more effectively learn global dependencies and periodic patterns. To integrate information from both time and frequency domains, we propose the $\textbf{T}$ime-$\textbf{F}$requency KAN (TFKAN). TFKAN employs a dual-branch architecture that independently processes features from each domain, ensuring that the distinct characteristics of each domain are fully utilized without interference. Additionally, to account for the heterogeneity between domains, we introduce a dimension-adjustment strategy that selectively upscales only in the frequency domain, enhancing efficiency while capturing richer frequency information. Experimental results demonstrate that TFKAN consistently outperforms state-of-the-art (SOTA) methods across multiple datasets. The code is available at https://github.com/LcWave/TFKAN.

Yao-Yuan Yang, Moto Hira, Zhaoheng Ni et al.

This document describes version 0.10 of TorchAudio: building blocks for machine learning applications in the audio and speech processing domain. The objective of TorchAudio is to accelerate the development and deployment of machine learning applications for researchers and engineers by providing off-the-shelf building blocks. The building blocks are designed to be GPU-compatible, automatically differentiable, and production-ready. TorchAudio can be easily installed from Python Package Index repository and the source code is publicly available under a BSD-2-Clause License (as of September 2021) at https://github.com/pytorch/audio. In this document, we provide an overview of the design principles, functionalities, and benchmarks of TorchAudio. We also benchmark our implementation of several audio and speech operations and models. We verify through the benchmarks that our implementations of various operations and models are valid and perform similarly to other publicly available implementations.

Yangyang Shi, Mei-Yuh Hwang, Xin Lei

Long Short Term Memory Connectionist Temporal Classification (LSTM-CTC) based end-to-end models are widely used in speech recognition due to its simplicity in training and efficiency in decoding. In conventional LSTM-CTC based models, a bottleneck projection matrix maps the hidden feature vectors obtained from LSTM to softmax output layer. In this paper, we propose to use a high rank projection layer to replace the projection matrix. The output from the high rank projection layer is a weighted combination of vectors that are projected from the hidden feature vectors via different projection matrices and non-linear activation function. The high rank projection layer is able to improve the expressiveness of LSTM-CTC models. The experimental results show that on Wall Street Journal (WSJ) corpus and LibriSpeech data set, the proposed method achieves 4%-6% relative word error rate (WER) reduction over the baseline CTC system. They outperform other published CTC based end-to-end (E2E) models under the condition that no external data or data augmentation is applied. Code has been made available at https://github.com/mobvoi/lstm_ctc.

Mingchen Zhuge, Changsheng Zhao, Dylan Ashley et al.

Contemporary evaluation techniques are inadequate for agentic systems. These approaches either focus exclusively on final outcomes -- ignoring the step-by-step nature of agentic systems, or require excessive manual labour. To address this, we introduce the Agent-as-a-Judge framework, wherein agentic systems are used to evaluate agentic systems. This is an organic extension of the LLM-as-a-Judge framework, incorporating agentic features that enable intermediate feedback for the entire task-solving process. We apply the Agent-as-a-Judge to the task of code generation. To overcome issues with existing benchmarks and provide a proof-of-concept testbed for Agent-as-a-Judge, we present DevAI, a new benchmark of 55 realistic automated AI development tasks. It includes rich manual annotations, like a total of 365 hierarchical user requirements. We benchmark three of the popular agentic systems using Agent-as-a-Judge and find it dramatically outperforms LLM-as-a-Judge and is as reliable as our human evaluation baseline. Altogether, we believe that Agent-as-a-Judge marks a concrete step forward for modern agentic systems -- by providing rich and reliable reward signals necessary for dynamic and scalable self-improvement.

Yang Li, Yuan Shangguan, Yuhao Wang et al.

Power consumption plays a crucial role in on-device streaming speech recognition, significantly influencing the user experience. This study explores how the configuration of weight parameters in speech recognition models affects their overall energy efficiency. We found that the influence of these parameters on power consumption varies depending on factors such as invocation frequency and memory allocation. Leveraging these insights, we propose design principles that enhance on-device speech recognition models by reducing power consumption with minimal impact on accuracy. Our approach, which adjusts model components based on their specific energy sensitivities, achieves up to 47% lower energy usage while preserving comparable model accuracy and improving real-time performance compared to leading methods.

Yangyang Shi, Qianqian Ren, Yong Liu et al.

Time series forecasting is crucial in many fields, yet current deep learning models struggle with noise, data sparsity, and capturing complex multi-scale patterns. This paper presents MFF-FTNet, a novel framework addressing these challenges by combining contrastive learning with multi-scale feature extraction across both frequency and time domains. MFF-FTNet introduces an adaptive noise augmentation strategy that adjusts scaling and shifting factors based on the statistical properties of the original time series data, enhancing model resilience to noise. The architecture is built around two complementary modules: a Frequency-Aware Contrastive Module (FACM) that refines spectral representations through frequency selection and contrastive learning, and a Complementary Time Domain Contrastive Module (CTCM) that captures both short- and long-term dependencies using multi-scale convolutions and feature fusion. A unified feature representation strategy enables robust contrastive learning across domains, creating an enriched framework for accurate forecasting. Extensive experiments on five real-world datasets demonstrate that MFF-FTNet significantly outperforms state-of-the-art models, achieving a 7.7% MSE improvement on multivariate tasks. These findings underscore MFF-FTNet's effectiveness in modeling complex temporal patterns and managing noise and sparsity, providing a comprehensive solution for both long- and short-term forecasting.

Frank Seide, Morrie Doulaty, Yangyang Shi et al.

We introduce Speech ReaLLM, a new ASR architecture that marries "decoder-only" ASR with the RNN-T to make multimodal LLM architectures capable of real-time streaming. This is the first "decoder-only" ASR architecture designed to handle continuous audio without explicit end-pointing. Speech ReaLLM is a special case of the more general ReaLLM ("real-time LLM") approach, also introduced here for the first time. The idea is inspired by RNN-T: Instead of generating a response only at the end of a user prompt, generate after every input token received in real time (it is often empty). On Librispeech "test", an 80M Speech ReaLLM achieves WERs of 3.0% and 7.4% in real time (without an external LM or auxiliary loss). This is only slightly above a 3x larger Attention-Encoder-Decoder baseline. We also show that this way, an LLM architecture can learn to represent and reproduce the flow of time; and that a pre-trained 7B LLM can be fine-tuned to do reasonably well on this task.

Zechun Liu, Barlas Oguz, Changsheng Zhao et al.

Several post-training quantization methods have been applied to large language models (LLMs), and have been shown to perform well down to 8-bits. We find that these methods break down at lower bit precision, and investigate quantization aware training for LLMs (LLM-QAT) to push quantization levels even further. We propose a data-free distillation method that leverages generations produced by the pre-trained model, which better preserves the original output distribution and allows quantizing any generative model independent of its training data, similar to post-training quantization methods. In addition to quantizing weights and activations, we also quantize the KV cache, which is critical for increasing throughput and support long sequence dependencies at current model sizes. We experiment with LLaMA models of sizes 7B, 13B, and 30B, at quantization levels down to 4-bits. We observe large improvements over training-free methods, especially in the low-bit settings.

Yassir Fathullah, Chunyang Wu, Yuan Shangguan et al.

State space models (SSMs) have recently shown promising results on small-scale sequence and language modelling tasks, rivalling and outperforming many attention-based approaches. In this paper, we propose a multi-head state space (MH-SSM) architecture equipped with special gating mechanisms, where parallel heads are taught to learn local and global temporal dynamics on sequence data. As a drop-in replacement for multi-head attention in transformer encoders, this new model significantly outperforms the transformer transducer on the LibriSpeech speech recognition corpus. Furthermore, we augment the transformer block with MH-SSMs layers, referred to as the Stateformer, achieving state-of-the-art performance on the LibriSpeech task, with word error rates of 1.76\%/4.37\% on the development and 1.91\%/4.36\% on the test sets without using an external language model.

Jay Mahadeokar, Yangyang Shi, Ke Li et al.

Streaming ASR with strict latency constraints is required in many speech recognition applications. In order to achieve the required latency, streaming ASR models sacrifice accuracy compared to non-streaming ASR models due to lack of future input context. Previous research has shown that streaming and non-streaming ASR for RNN Transducers can be unified by cascading causal and non-causal encoders. This work improves upon this cascaded encoders framework by leveraging two streaming non-causal encoders with variable input context sizes that can produce outputs at different audio intervals (e.g. fast and slow). We propose a novel parallel time-synchronous beam search algorithm for transducers that decodes from fast-slow encoders, where the slow encoder corrects the mistakes generated from the fast encoder. The proposed algorithm, achieves up to 20% WER reduction with a slight increase in token emission delays on the public Librispeech dataset and in-house datasets. We also explore techniques to reduce the computation by distributing processing between the fast and slow encoders. Lastly, we explore sharing the parameters in the fast encoder to reduce the memory footprint. This enables low latency processing on edge devices with low computation cost and a low memory footprint.

Yangyang Shi, Chunyang Wu, Dilin Wang et al.

This paper improves the streaming transformer transducer for speech recognition by using non-causal convolution. Many works apply the causal convolution to improve streaming transformer ignoring the lookahead context. We propose to use non-causal convolution to process the center block and lookahead context separately. This method leverages the lookahead context in convolution and maintains similar training and decoding efficiency. Given the similar latency, using the non-causal convolution with lookahead context gives better accuracy than causal convolution, especially for open-domain dictation scenarios. Besides, this paper applies talking-head attention and a novel history context compression scheme to further improve the performance. The talking-head attention improves the multi-head self-attention by transferring information among different heads. The history context compression method introduces more extended history context compactly. On our in-house data, the proposed methods improve a small Emformer baseline with lookahead context by relative WERR 5.1\%, 14.5\%, 8.4\% on open-domain dictation, assistant general scenarios, and assistant calling scenarios, respectively.

Dawei Liang, Yangyang Shi, Yun Wang et al.

Detection of common events and scenes from audio is useful for extracting and understanding human contexts in daily life. Prior studies have shown that leveraging knowledge from a relevant domain is beneficial for a target acoustic event detection (AED) process. Inspired by the observation that many human-centered acoustic events in daily life involve voice elements, this paper investigates the potential of transferring high-level voice representations extracted from a public speaker dataset to enrich an AED pipeline. Towards this end, we develop a dual-branch neural network architecture for the joint learning of voice and acoustic features during an AED process and conduct thorough empirical studies to examine the performance on the public AudioSet [1] with different types of inputs. Our main observations are that: 1) Joint learning of audio and voice inputs improves the AED performance (mean average precision) for both a CNN baseline (0.292 vs 0.134 mAP) and a TALNet [2] baseline (0.361 vs 0.351 mAP); 2) Augmenting the extra voice features is critical to maximize the model performance with dual inputs.

Xiaohui Zhang, Vimal Manohar, David Zhang et al.

Hybrid automatic speech recognition (ASR) models are typically sequentially trained with CTC or LF-MMI criteria. However, they have vastly different legacies and are usually implemented in different frameworks. In this paper, by decoupling the concepts of modeling units and label topologies and building proper numerator/denominator graphs accordingly, we establish a generalized framework for hybrid acoustic modeling (AM). In this framework, we show that LF-MMI is a powerful training criterion applicable to both limited-context and full-context models, for wordpiece/mono-char/bi-char/chenone units, with both HMM/CTC topologies. From this framework, we propose three novel training schemes: chenone(ch)/wordpiece(wp)-CTC-bMMI, and wordpiece(wp)-HMM-bMMI with different advantages in training performance, decoding efficiency and decoding time-stamp accuracy. The advantages of different training schemes are evaluated comprehensively on Librispeech, and wp-CTC-bMMI and ch-CTC-bMMI are evaluated on two real world ASR tasks to show their effectiveness. Besides, we also show bi-char(bc) HMM-MMI models can serve as better alignment models than traditional non-neural GMM-HMMs.

Varun Nagaraja, Yangyang Shi, Ganesh Venkatesh et al.

On-device speech recognition requires training models of different sizes for deploying on devices with various computational budgets. When building such different models, we can benefit from training them jointly to take advantage of the knowledge shared between them. Joint training is also efficient since it reduces the redundancy in the training procedure's data handling operations. We propose a method for collaboratively training acoustic encoders of different sizes for speech recognition. We use a sequence transducer setup where different acoustic encoders share a common predictor and joiner modules. The acoustic encoders are also trained using co-distillation through an auxiliary task for frame level chenone prediction, along with the transducer loss. We perform experiments using the LibriSpeech corpus and demonstrate that the collaboratively trained acoustic encoders can provide up to a 11% relative improvement in the word error rate on both the test partitions.

Jay Mahadeokar, Yangyang Shi, Yuan Shangguan et al.

Often, the storage and computational constraints of embeddeddevices demand that a single on-device ASR model serve multiple use-cases / domains. In this paper, we propose aFlexibleTransducer(FlexiT) for on-device automatic speech recognition to flexibly deal with multiple use-cases / domains with different accuracy and latency requirements. Specifically, using a single compact model, FlexiT provides a fast response for voice commands, and accurate transcription but with more latency for dictation. In order to achieve flexible and better accuracy and latency trade-offs, the following techniques are used. Firstly, we propose using domain-specific altering of segment size for Emformer encoder that enables FlexiT to achieve flexible de-coding. Secondly, we use Alignment Restricted RNNT loss to achieve flexible fine-grained control on token emission latency for different domains. Finally, we add a domain indicator vector as an additional input to the FlexiT model. Using the combination of techniques, we show that a single model can be used to improve WERs and real time factor for dictation scenarios while maintaining optimal latency for voice commands use-cases

Yuan Shangguan, Rohit Prabhavalkar, Hang Su et al.

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

Duc Le, Mahaveer Jain, Gil Keren et al.

How to leverage dynamic contextual information in end-to-end speech recognition has remained an active research area. Previous solutions to this problem were either designed for specialized use cases that did not generalize well to open-domain scenarios, did not scale to large biasing lists, or underperformed on rare long-tail words. We address these limitations by proposing a novel solution that combines shallow fusion, trie-based deep biasing, and neural network language model contextualization. These techniques result in significant 19.5% relative Word Error Rate improvement over existing contextual biasing approaches and 5.4%-9.3% improvement compared to a strong hybrid baseline on both open-domain and constrained contextualization tasks, where the targets consist of mostly rare long-tail words. Our final system remains lightweight and modular, allowing for quick modification without model re-training.

Yangyang Shi, Varun Nagaraja, Chunyang Wu et al.

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

Ching-Feng Yeh, Yongqiang Wang, Yangyang Shi et al.

Attention-based models have been gaining popularity recently for their strong performance demonstrated in fields such as machine translation and automatic speech recognition. One major challenge of attention-based models is the need of access to the full sequence and the quadratically growing computational cost concerning the sequence length. These characteristics pose challenges, especially for low-latency scenarios, where the system is often required to be streaming. In this paper, we build a compact and streaming speech recognition system on top of the end-to-end neural transducer architecture with attention-based modules augmented with convolution. The proposed system equips the end-to-end models with the streaming capability and reduces the large footprint from the streaming attention-based model using augmented memory. On the LibriSpeech dataset, our proposed system achieves word error rates 2.7% on test-clean and 5.8% on test-other, to our best knowledge the lowest among streaming approaches reported so far.

Yongqiang Wang, Yangyang Shi, Frank Zhang et al.

In this paper, we summarize the application of transformer and its streamable variant, Emformer based acoustic model for large scale speech recognition applications. We compare the transformer based acoustic models with their LSTM counterparts on industrial scale tasks. Specifically, we compare Emformer with latency-controlled BLSTM (LCBLSTM) on medium latency tasks and LSTM on low latency tasks. On a low latency voice assistant task, Emformer gets 24% to 26% relative word error rate reductions (WERRs). For medium latency scenarios, comparing with LCBLSTM with similar model size and latency, Emformer gets significant WERR across four languages in video captioning datasets with 2-3 times inference real-time factors reduction.

Yangyang Shi, Yongqiang Wang, Chunyang Wu et al.

This paper proposes an efficient memory transformer Emformer for low latency streaming speech recognition. In Emformer, the long-range history context is distilled into an augmented memory bank to reduce self-attention's computation complexity. A cache mechanism saves the computation for the key and value in self-attention for the left context. Emformer applies a parallelized block processing in training to support low latency models. We carry out experiments on benchmark LibriSpeech data. Under average latency of 960 ms, Emformer gets WER $2.50\%$ on test-clean and $5.62\%$ on test-other. Comparing with a strong baseline augmented memory transformer (AM-TRF), Emformer gets $4.6$ folds training speedup and $18\%$ relative real-time factor (RTF) reduction in decoding with relative WER reduction $17\%$ on test-clean and $9\%$ on test-other. For a low latency scenario with an average latency of 80 ms, Emformer achieves WER $3.01\%$ on test-clean and $7.09\%$ on test-other. Comparing with the LSTM baseline with the same latency and model size, Emformer gets relative WER reduction $9\%$ and $16\%$ on test-clean and test-other, respectively.

Yangyang Shi, Yongqiang Wang, Chunyang Wu et al.

Transformers, originally proposed for natural language processing (NLP) tasks, have recently achieved great success in automatic speech recognition (ASR). However, adjacent acoustic units (i.e., frames) are highly correlated, and long-distance dependencies between them are weak, unlike text units. It suggests that ASR will likely benefit from sparse and localized attention. In this paper, we propose Weak-Attention Suppression (WAS), a method that dynamically induces sparsity in attention probabilities. We demonstrate that WAS leads to consistent Word Error Rate (WER) improvement over strong transformer baselines. On the widely used LibriSpeech benchmark, our proposed method reduced WER by 10%$ on test-clean and 5% on test-other for streamable transformers, resulting in a new state-of-the-art among streaming models. Further analysis shows that WAS learns to suppress attention of non-critical and redundant continuous acoustic frames, and is more likely to suppress past frames rather than future ones. It indicates the importance of lookahead in attention-based ASR models.

Chunyang Wu, Yongqiang Wang, Yangyang Shi et al.

Transformer-based acoustic modeling has achieved great suc-cess for both hybrid and sequence-to-sequence speech recogni-tion. However, it requires access to the full sequence, and thecomputational cost grows quadratically with respect to the in-put sequence length. These factors limit its adoption for stream-ing applications. In this work, we proposed a novel augmentedmemory self-attention, which attends on a short segment of theinput sequence and a bank of memories. The memory bankstores the embedding information for all the processed seg-ments. On the librispeech benchmark, our proposed methodoutperforms all the existing streamable transformer methods bya large margin and achieved over 15% relative error reduction,compared with the widely used LC-BLSTM baseline. Our find-ings are also confirmed on some large internal datasets.

Yangyang Shi, Mei-Yuh Hwang, Xin Lei et al.

Recurrent Neural Networks (RNNs) have dominated language modeling because of their superior performance over traditional N-gram based models. In many applications, a large Recurrent Neural Network language model (RNNLM) or an ensemble of several RNNLMs is used. These models have large memory footprints and require heavy computation. In this paper, we examine the effect of applying knowledge distillation in reducing the model size for RNNLMs. In addition, we propose a trust regularization method to improve the knowledge distillation training for RNNLMs. Using knowledge distillation with trust regularization, we reduce the parameter size to a third of that of the previously published best model while maintaining the state-of-the-art perplexity result on Penn Treebank data. In a speech recognition N-bestrescoring task, we reduce the RNNLM model size to 18.5% of the baseline system, with no degradation in word error rate(WER) performance on Wall Street Journal data set.