Large language models (LLMs) have proven themselves highly flexible, able to solve a wide range of generative tasks, such as abstractive summarization and open-ended question answering. In this paper we extend the capabilities of LLM by directly attaching a small audio encoder allowing it to perform speech recognition. By directly prepending a sequence of audio embeddings to the text token embeddings, the LLM can be converted to an automatic speech recognition (ASR) system, and be used in the exact same manner as its textual counterpart. Experiments on Multilingual LibriSpeech (MLS) show that incorporating a conformer encoder into the open sourced LLaMA-7B allows it to outperform monolingual baselines by 18% relatively in WER and perform multilingual speech recognition, despite LLaMA being trained overwhelmingly on English text. Furthermore, we perform ablation studies to investigate whether the LLM can be completely frozen during training to maintain its original capabilities, scaling up the audio encoder, and increasing the audio encoder striding to generate fewer embeddings. The results from these studies show that multilingual ASR is possible even when the LLM is frozen, or when strides of almost 1 second are used in the audio encoder opening up the possibility for LLMs to operate on long-form audio.

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.

Automatic Speech Recognition (ASR) models need to be optimized for specific hardware before they can be deployed on devices. This can be done by tuning the model’s hyperparameters or exploring variations in its architecture. Re-training and re-validating models after making these changes can be a resource-intensive task. This paper presents TODM (Train Once Deploy Many), a new approach to efficiently train many sizes of hardware-friendly on-device ASR models with comparable GPU-hours to that of a single training job. TODM leverages insights from prior work on Supernet, where Recurrent Neural Network Transducer (RNN-T) models share weights within a Supernet. It reduces layer sizes and widths of the Supernet to obtain subnetworks, making them smaller models suitable for all hardware types. We introduce a novel combination of three techniques to improve the outcomes of the TODM Supernet: adaptive dropout, an in-place Alpha-divergence knowledge distillation, and the use of ScaledAdam optimizer. We validate our approach by comparing Supernet-trained versus individually tuned Multi-Head State Space Model (MH-SSM) RNN-T using LibriSpeech. Results demonstrate that our TODM Supernet either matches or surpasses the performance of manually tuned models by up to a relative of 3% better in word error rate (WER), while efficiently keeping the cost of training many models at a small constant.

Neural network pruning offers an effective method for compressing a multilingual automatic speech recognition (ASR) model with minimal performance loss. However, it entails several rounds of pruning and re-training needed to be run for each language. In this work, we propose the use of an adaptive masking approach in two scenarios for pruning a multilingual ASR model efficiently, each resulting in sparse monolingual models or a sparse multilingual model (named as Dynamic ASR Pathways). Our approach dynamically adapts the subnetwork, avoiding premature decisions about a fixed sub-network structure. We show that our approach outperforms existing pruning methods when targeting sparse monolingual models. Further, we illustrate that Dynamic ASR Pathways jointly discovers and trains better sub-networks (pathways) of a single multilingual model by adapting from different sub-network initializations, thereby reducing the need for language-specific pruning.

Yassir Fathullah, Chunyang Wu, Egor Lakomkin, Ke Li, Junteng Jia, Yuan Shangguan, Jay Mahadeokar, Ozlem Kalinli, Christian Fuegen, Mike Seltzer. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

This paper improves the streaming transformer transducer for speech recognition 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. 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.

From wearables to powerful smart devices, modern automatic speech recognition (ASR) models run on a variety of edge devices with different computational budgets. To navigate the Pareto front of model accuracy vs model size, researchers are trapped in a dilemma of optimizing model accuracy by training and fine-tuning models for each individual edge device while keeping the training GPU-hours tractable. In this paper, we propose Omni-sparsity DNN, where a single neural network can be pruned to generate optimized model for a large range of model sizes. We develop training strategies for Omni-sparsity DNN that allows it to find models along the Pareto front of word-error-rate (WER) vs model size while keeping the training GPU-hours to no more than that of training one singular model. We demonstrate the Omni-sparsity DNN with streaming E2E ASR models. Our results show great saving on training time and resources with similar or better accuracy on LibriSpeech compared to individually pruned sparse models: 2%-6.6% better WER on Test-other.

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.

Recurrent Neural Network Transducer (RNN-T), like most end-to-end speech recognition model architectures, has an implicit neural network language model (NNLM) and cannot easily leverage unpaired text data during training. Previous work has proposed various fusion methods to incorporate external NNLMs into end-to-end ASR to address this weakness. In this paper, we propose extensions to these techniques that allow RNN-T to exploit external NNLMs during both training and inference time, resulting in 13-18% relative Word Error Rate improvement on Librispeech compared to strong baselines. Furthermore, our methods do not incur extra algorithmic latency and allow for flexible plug- and-play of different NNLMs without re-training. We also share in-depth analysis to better understand the benefits of the different NNLM fusion methods. Our work provides a reliable technique for leveraging unpaired text data to significantly improve RNN-T while keeping the system streamable, flexible, and lightweight.

Knowledge Distillation is an effective method of transferring knowledge from a large model to a smaller model. Distillation can be viewed as a type of model compression, and has played an important role for on-device ASR applications. In this paper, we develop a distillation method for RNN-Transducer (RNN-T) models, a popular end-to-end neural network architecture for streaming speech recognition. Our proposed distillation loss is simple and efficient, and uses only the "y" and "blank" posterior probabilities from the RNN-T output probability lattice. We study the effectiveness of the proposed approach in improving the accuracy of sparse RNN-T models obtained by gradually pruning a larger uncompressed model, which also serves as the teacher during distillation. With distillation of 60% and 90% sparse multi-domain RNN-T models, we obtain WER reductions of 4.3% and 12.1% respectively, on a noisy FarField eval set. We also present results of experiments on LibriSpeech, where the introduction of the distillation loss yields a 4.8% relative WER reduction on the test-other dataset for a small Conformer model.

Recurrent transducer models have emerged as a promising solution for speech recognition on the current and next generation smart devices. The transducer models provide competitive accuracy within a reasonable memory footprint alleviating the memory capacity constraints in these devices. However, these models access parameters from off-chip memory for every input time step which adversely effects device battery life and limits their usability on low-power devices.We address transducer model’s memory access concerns by optimizing their model architecture and designing novel recurrent cell designs. We demonstrate that i) model’s energy cost is dominated by accessing model weights from off-chip memory, ii) transducer model architecture is pivotal in determining the number of accesses to off-chip memory and just model size is not a good proxy, iii) our transducer model optimizations and novel recurrent cell reduces off-chip memory accesses by 4.5× and model size by 2× with minimal accuracy impact.

How to leverage dynamic contextual information in end-toend 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.

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.

There is a growing interest in the speech community in developing Recurrent Neural Network Transducer (RNN-T) models for automatic speech recognition (ASR) applications. RNN-T is trained with a loss function that does not enforce temporal alignment of the training transcripts and audio. As a result, RNN-T models built with uni-directional long short term memory (LSTM) encoders tend to wait for longer spans of input audio, before streaming already decoded ASR tokens. In this work, we propose a modification to the RNN-T loss function and develop Alignment Restricted RNN-T (Ar-RNN-T) models, which utilize audio-text alignment in-formation to guide the loss computation. We compare the proposed method with existing works, such as monotonic RNN-T, on LibriSpeech and in-house datasets. We show that the Ar-RNN-T loss provides a refined control to navigate the trade-offs between the token emission delays and the Word Error Rate (WER). The Ar-RNN-T models also improve downstream applications such as the ASR End-pointing by guaranteeing token emissions within any given range of latency. Moreover, the Ar-RNN-T loss allows for bigger batch sizes and 4 times higher throughput for our LSTM model architecture, enabling faster training and convergence on GPUs.

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains [1] to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

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

End-to-end (E2E) models, which directly predict output character sequences given input speech, are good candidates for on-device speech recognition. E2E models, however, present numerous challenges: In order to be truly useful, such models must decode speech utterances in a streaming fashion, in real time; they must be robust to the long tail of use cases; they must be able to leverage user-specific context (e.g., contact lists); and above all, they must be extremely accurate. In this work, we describe our efforts at building an E2E speech recog-nizer using a recurrent neural network transducer. In experimental evaluations, we find that the proposed approach can outperform a conventional CTC-based model in terms of both latency and accuracy in a number of evaluation categories.

Automatic recognition of emotion in speech is an active area of research. One of the important open challenges relates to how the emotional characteristics of speech change in time. Past research has demonstrated the importance of capturing global dynamics (across an entire utterance) and local dynamics (within segments of an utterance). In this paper, we propose a novel concept, EmoShapelets, to capture the local dynamics in speech. EmoShapelets capture changes in emotion that occur within utterances. We propose a framework to generate, update, and select EmoShapelets. We also demonstrate the discriminative power of EmoShapelets by using them with various classifiers to achieve comparable results with the state-of-the-art systems on the IEMOCAP dataset. EmoShapelets can serve as basic units of emotion expression and provide additional evidence supporting the existence of local patterns of emotion underlying human communication.

Emotion is central to communication; it colors our interpretation of events and social interactions. Emotion expression is generally multimodal, modulating our facial movement, vocal behavior, and body gestures. The method through which this multimodal information is integrated and perceived is not well understood. This knowledge has implications for the design of multimodal classification algorithms, affective interfaces, and even mental health assessment. We present a novel data set designed to support research into the emotion perception process, the University of Michigan Emotional McGurk Effect Data set (UMEME). UMEME has a critical feature that differentiates it from currently existing data sets; it contains not only emotionally congruent stimuli (emotionally matched faces and voices), but also emotionally incongruent stimuli (emotionally mismatched faces and voices). The inclusion of emotionally complex and dynamic stimuli provides an opportunity to study how individuals make assessments of emotion content in the presence of emotional incongruence, or emotional noise. We describe the collection, annotation, and statistical properties of the data and present evidence illustrating how audio and video interact to result in specific types of emotion perception. The results demonstrate that there exist consistent patterns underlying emotion evaluation, even given incongruence, positioning UMEME as an important new tool for understanding emotion perception.