Ozlem Kalinli

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Yassir Fathullah, Chunyang Wu, Egor Lakomkin et al.

Large language models 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 LLMs by directly attaching a small audio encoder allowing it to perform speech recognition. By directly prepending a sequence of audial 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% 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.

Irina-Elena Veliche, Zhuangqun Huang, Vineeth Ayyat Kochaniyan et al.

The current public datasets for speech recognition (ASR) tend not to focus specifically on the fairness aspect, such as performance across different demographic groups. This paper introduces a novel dataset, Fair-Speech, a publicly released corpus to help researchers evaluate their ASR models for accuracy across a diverse set of self-reported demographic information, such as age, gender, ethnicity, geographic variation and whether the participants consider themselves native English speakers. Our dataset includes approximately 26.5K utterances in recorded speech by 593 people in the United States, who were paid to record and submit audios of themselves saying voice commands. We also provide ASR baselines, including on models trained on transcribed and untranscribed social media videos and open source models.

Duc Le, Akshat Shrivastava, Paden Tomasello et al. · meta-ai

We propose a novel deliberation-based approach to end-to-end (E2E) spoken language understanding (SLU), where a streaming automatic speech recognition (ASR) model produces the first-pass hypothesis and a second-pass natural language understanding (NLU) component generates the semantic parse by conditioning on both ASR's text and audio embeddings. By formulating E2E SLU as a generalized decoder, our system is able to support complex compositional semantic structures. Furthermore, the sharing of parameters between ASR and NLU makes the system especially suitable for resource-constrained (on-device) environments; our proposed approach consistently outperforms strong pipeline NLU baselines by 0.60% to 0.65% on the spoken version of the TOPv2 dataset (STOP). We demonstrate that the fusion of text and audio features, coupled with the system's ability to rewrite the first-pass hypothesis, makes our approach more robust to ASR errors. Finally, we show that our approach can significantly reduce the degradation when moving from natural speech to synthetic speech training, but more work is required to make text-to-speech (TTS) a viable solution for scaling up E2E SLU.

Suyoun Kim, Akshat Shrivastava, Duc Le et al. · meta-ai

End-to-end (E2E) spoken language understanding (SLU) systems that generate a semantic parse from speech have become more promising recently. This approach uses a single model that utilizes audio and text representations from pre-trained speech recognition models (ASR), and outperforms traditional pipeline SLU systems in on-device streaming scenarios. However, E2E SLU systems still show weakness when text representation quality is low due to ASR transcription errors. To overcome this issue, we propose a novel E2E SLU system that enhances robustness to ASR errors by fusing audio and text representations based on the estimated modality confidence of ASR hypotheses. We introduce two novel techniques: 1) an effective method to encode the quality of ASR hypotheses and 2) an effective approach to integrate them into E2E SLU models. We show accuracy improvements on STOP dataset and share the analysis to demonstrate the effectiveness of our approach.

Egor Lakomkin, Chunyang Wu, Yassir Fathullah et al.

In recent years, Large Language Models (LLMs) have garnered significant attention from the research community due to their exceptional performance and generalization capabilities. In this paper, we introduce a novel method for contextualizing speech recognition models incorporating LLMs. Our approach casts speech recognition as a mixed-modal language modeling task based on a pretrained LLM. We provide audio features, along with optional text tokens for context, to train the system to complete transcriptions in a decoder-only fashion. As a result, the system is implicitly incentivized to learn how to leverage unstructured contextual information during training. Our empirical results demonstrate a significant improvement in performance, with a 6% WER reduction when additional textual context is provided. Moreover, we find that our method performs competitively and improve by 7.5% WER overall and 17% WER on rare words against a baseline contextualized RNN-T system that has been trained on more than twenty five times larger speech dataset. Overall, we demonstrate that by only adding a handful number of trainable parameters via adapters, we can unlock contextualized speech recognition capability for the pretrained LLM while keeping the same text-only input functionality.

Roshan Sharma, Suyoun Kim, Daniel Lazar et al. · meta-ai

Spoken semantic parsing (SSP) involves generating machine-comprehensible parses from input speech. Training robust models for existing application domains represented in training data or extending to new domains requires corresponding triplets of speech-transcript-semantic parse data, which is expensive to obtain. In this paper, we address this challenge by examining methods that can use transcript-semantic parse data (unpaired text) without corresponding speech. First, when unpaired text is drawn from existing textual corpora, Joint Audio Text (JAT) and Text-to-Speech (TTS) are compared as ways to generate speech representations for unpaired text. Experiments on the STOP dataset show that unpaired text from existing and new domains improves performance by 2% and 30% in absolute Exact Match (EM) respectively. Second, we consider the setting when unpaired text is not available in existing textual corpora. We propose to prompt Large Language Models (LLMs) to generate unpaired text for existing and new domains. Experiments show that examples and words that co-occur with intents can be used to generate unpaired text with Llama 2.0. Using the generated text with JAT and TTS for spoken semantic parsing improves EM on STOP by 1.4% and 2.6% absolute for existing and new domains respectively.

Mu Yang, Andros Tjandra, Chunxi Liu et al.

Neural network pruning compresses automatic speech recognition (ASR) models effectively. However, in multilingual ASR, language-agnostic pruning may lead to severe performance drops on some languages because language-agnostic pruning masks may not fit all languages and discard important language-specific parameters. In this work, we present ASR pathways, a sparse multilingual ASR model that activates language-specific sub-networks ("pathways"), such that the parameters for each language are learned explicitly. With the overlapping sub-networks, the shared parameters can also enable knowledge transfer for lower-resource languages via joint multilingual training. We propose a novel algorithm to learn ASR pathways, and evaluate the proposed method on 4 languages with a streaming RNN-T model. Our proposed ASR pathways outperform both dense models and a language-agnostically pruned model, and provide better performance on low-resource languages compared to the monolingual sparse models.

Andros Tjandra, Nayan Singhal, David Zhang et al.

End-to-end multilingual ASR has become more appealing because of several reasons such as simplifying the training and deployment process and positive performance transfer from high-resource to low-resource languages. However, scaling up the number of languages, total hours, and number of unique tokens is not a trivial task. This paper explores large-scale multilingual ASR models on 70 languages. We inspect two architectures: (1) Shared embedding and output and (2) Multiple embedding and output model. In the shared model experiments, we show the importance of tokenization strategy across different languages. Later, we use our optimal tokenization strategy to train multiple embedding and output model to further improve our result. Our multilingual ASR achieves 13.9%-15.6% average WER relative improvement compared to monolingual models. We show that our multilingual ASR generalizes well on an unseen dataset and domain, achieving 9.5% and 7.5% WER on Multilingual Librispeech (MLS) with zero-shot and finetuning, respectively.

Yassir Fathullah, Chunyang Wu, Egor Lakomkin et al.

In this work, we extend the instruction-tuned Llama-2 model with end-to-end general-purpose speech processing and reasoning abilities while maintaining the wide range of original LLM capabilities, without using any carefully curated paired data. The resulting end-to-end model, named AudioChatLlama, can utilize audio prompts as a replacement for text and sustain a conversation. Such a model also has extended cross-modal capabilities such as being able to perform spoken question answering (QA), speech translation, and audio summarization amongst many other closed and open-domain tasks. This is unlike prior approaches in speech, in which LLMs are extended to handle audio for a limited number of pre-designated tasks. On both synthesized and recorded speech QA test sets, evaluations show that our end-to-end approach is on par with or outperforms cascaded systems (speech recognizer + LLM) in terms of modeling the response to a prompt. Furthermore, unlike cascades, our approach can interchange text and audio modalities and intrinsically utilize prior context in a conversation to provide better results.

Sung-Lin Yeh, Wei Zhou, Gil Keren et al.

Recent speech language models rely on encoders that are optimized separately from autoregressive models. Since these encoders are unaware of the downstream objectives, the extracted representations may not be optimal for downstream tasks. To address this limitation, we introduce a discrete latent variable model on mel spectrograms that jointly optimizes the encoder and the speech language model. Joint optimization not only brings improvements over codec-based and other mel-spectrogram-based baselines on zero-shot Text-to-Speech (TTS) and Speech-to-Text (STT) tasks, but also effectively alleviates common issues in autoregressive mel-spectrogram modeling, such as prolonged silence generation and word omissions.

Suyoun Kim, Ke Li, Lucas Kabela et al.

Recently, there has been an increasing interest in two-pass streaming end-to-end speech recognition (ASR) that incorporates a 2nd-pass rescoring model on top of the conventional 1st-pass streaming ASR model to improve recognition accuracy while keeping latency low. One of the latest 2nd-pass rescoring model, Transformer Rescorer, takes the n-best initial outputs and audio embeddings from the 1st-pass model, and then choose the best output by re-scoring the n-best initial outputs. However, training this Transformer Rescorer requires expensive paired audio-text training data because the model uses audio embeddings as input. In this work, we present our Joint Audio/Text training method for Transformer Rescorer, to leverage unpaired text-only data which is relatively cheaper than paired audio-text data. We evaluate Transformer Rescorer with our Joint Audio/Text training on Librispeech dataset as well as our large-scale in-house dataset and show that our training method can improve word error rate (WER) significantly compared to standard Transformer Rescorer without requiring any extra model parameters or latency.

Jiamin Xie, Ke Li, Jinxi Guo et al.

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 sub-network, 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.

Yuan Shangguan, Haichuan Yang, Danni Li et al.

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 dropouts, 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.

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.

Arpita Vats, Zhe Liu, Peng Su et al.

Model adaptation is crucial to handle the discrepancy between proxy training data and actual users data received. To effectively perform adaptation, textual data of users is typically stored on servers or their local devices, where downstream natural language processing (NLP) models can be directly trained using such in-domain data. However, this might raise privacy and security concerns due to the extra risks of exposing user information to adversaries. Replacing identifying information in textual data with a generic marker has been recently explored. In this work, we leverage large language models (LLMs) to suggest substitutes of masked tokens and have their effectiveness evaluated on downstream language modeling tasks. Specifically, we propose multiple pre-trained and fine-tuned LLM-based approaches and perform empirical studies on various datasets for the comparison of these methods. Experimental results show that models trained on the obfuscation corpora are able to achieve comparable performance with the ones trained on the original data without privacy-preserving token masking.

Yancheng Wang, Osama Hanna, Ruiming Xie et al.

Emotion recognition in speech presents a complex multimodal challenge, requiring comprehension of both linguistic content and vocal expressivity, particularly prosodic features such as fundamental frequency, intensity, and temporal dynamics. Although large language models (LLMs) have shown promise in reasoning over textual transcriptions for emotion recognition, they typically neglect fine-grained prosodic information, limiting their effectiveness and interpretability. In this work, we propose VowelPrompt, a linguistically grounded framework that augments LLM-based emotion recognition with interpretable, fine-grained vowel-level prosodic cues. Drawing on phonetic evidence that vowels serve as primary carriers of affective prosody, VowelPrompt extracts pitch-, energy-, and duration-based descriptors from time-aligned vowel segments, and converts these features into natural language descriptions for better interpretability. Such a design enables LLMs to jointly reason over lexical semantics and fine-grained prosodic variation. Moreover, we adopt a two-stage adaptation procedure comprising supervised fine-tuning (SFT) followed by Reinforcement Learning with Verifiable Reward (RLVR), implemented via Group Relative Policy Optimization (GRPO), to enhance reasoning capability, enforce structured output adherence, and improve generalization across domains and speaker variations. Extensive evaluations across diverse benchmark datasets demonstrate that VowelPrompt consistently outperforms state-of-the-art emotion recognition methods under zero-shot, fine-tuned, cross-domain, and cross-linguistic conditions, while enabling the generation of interpretable explanations that are jointly grounded in contextual semantics and fine-grained prosodic structure.

Zhe Liu, Ozlem Kalinli

Recent research has shown that language models have a tendency to memorize rare or unique token sequences in the training corpus. After deploying a model, practitioners might be asked to delete any personal information from the model by individuals' requests. Re-training the underlying model every time individuals would like to practice their rights to be forgotten is computationally expensive. We employ a teacher-student framework and propose a novel leave-one-out ensemble method to unlearn the targeted textual sequences that need to be forgotten from the model. In our approach, multiple teachers are trained on disjoint sets; for each targeted sequence to be removed, we exclude the teacher trained on the set containing this sequence and aggregate the predictions from remaining teachers to provide supervision during fine-tuning. Experiments on LibriSpeech and WikiText-103 datasets show that the proposed method achieves superior privacy-utility trade-offs than other counterparts.

Yingyi Ma, Zhe Liu, Ozlem Kalinli

Language models (LMs) have been commonly adopted to boost the performance of automatic speech recognition (ASR) particularly in domain adaptation tasks. Conventional way of LM training treats all the words in corpora equally, resulting in suboptimal improvements in ASR performance. In this work, we introduce a novel correction focused LM training approach which aims to prioritize ASR fallible words. The word-level ASR fallibility score, representing the likelihood of ASR mis-recognition, is defined and shaped as a prior word distribution to guide the LM training. To enable correction focused training with text-only corpora, large language models (LLMs) are employed as fallibility score predictors and text generators through multi-task fine-tuning. Experimental results for domain adaptation tasks demonstrate the effectiveness of our proposed method. Compared with conventional LMs, correction focused training achieves up to relatively 5.5% word error rate (WER) reduction in sufficient text scenarios. In insufficient text scenarios, LM training with LLM-generated text achieves up to relatively 13% WER reduction, while correction focused training further obtains up to relatively 6% WER reduction.

Jinxi Guo, Niko Moritz, Yingyi Ma et al.

The internal language model (ILM) of the neural transducer has been widely studied. In most prior work, it is mainly used for estimating the ILM score and is subsequently subtracted during inference to facilitate improved integration with external language models. Recently, various of factorized transducer models have been proposed, which explicitly embrace a standalone internal language model for non-blank token prediction. However, even with the adoption of factorized transducer models, limited improvement has been observed compared to shallow fusion. In this paper, we propose a novel ILM training and decoding strategy for factorized transducer models, which effectively combines the blank, acoustic and ILM scores. Our experiments show a 17% relative improvement over the standard decoding method when utilizing a well-trained ILM and the proposed decoding strategy on LibriSpeech datasets. Furthermore, when compared to a strong RNN-T baseline enhanced with external LM fusion, the proposed model yields a 5.5% relative improvement on general-sets and an 8.9% WER reduction for rare words. The proposed model can achieve superior performance without relying on external language models, rendering it highly efficient for production use-cases. To further improve the performance, we propose a novel and memory-efficient ILM-fusion-aware minimum word error rate (MWER) training method which improves ILM integration significantly.

Keqi Deng, Jinxi Guo, Yingyi Ma et al.

While large language models (LLMs) have been applied to automatic speech recognition (ASR), the task of making the model streamable remains a challenge. This paper proposes a novel model architecture, Transducer-Llama, that integrates LLMs into a Factorized Transducer (FT) model, naturally enabling streaming capabilities. Furthermore, given that the large vocabulary of LLMs can cause data sparsity issue and increased training costs for spoken language systems, this paper introduces an efficient vocabulary adaptation technique to align LLMs with speech system vocabularies. The results show that directly optimizing the FT model with a strong pre-trained LLM-based predictor using the RNN-T loss yields some but limited improvements over a smaller pre-trained LM predictor. Therefore, this paper proposes a weak-to-strong LM swap strategy, using a weak LM predictor during RNN-T loss training and then replacing it with a strong LLM. After LM replacement, the minimum word error rate (MWER) loss is employed to finetune the integration of the LLM predictor with the Transducer-Llama model. Experiments on the LibriSpeech and large-scale multi-lingual LibriSpeech corpora show that the proposed streaming Transducer-Llama approach gave a 17% relative WER reduction (WERR) over a strong FT baseline and a 32% WERR over an RNN-T baseline.

Wei Zhou, Junteng Jia, Leda Sari et al.

CTC compressor can be an effective approach to integrate audio encoders to decoder-only models, which has gained growing interest for different speech applications. In this work, we propose a novel CTC compressor based joint speech and text training (CJST) framework for decoder-only ASR. CJST matches speech and text modalities from both directions by exploring a simple modality adaptor and several features of the CTC compressor, including sequence compression, on-the-fly forced peaky alignment and CTC class embeddings. Experimental results on the Librispeech and TED-LIUM2 corpora show that the proposed CJST achieves an effective text injection without the need of duration handling, leading to the best performance for both in-domain and cross-domain scenarios. We also provide a comprehensive study on CTC compressor, covering various compression modes, edge case handling and behavior under both clean and noisy data conditions, which reveals the most robust setting to use CTC compressor for decoder-only models.

Yi-Jen Shih, Desh Raj, Chunyang Wu et al.

Recent advances in speech large language models (speech LLMs) have enabled seamless spoken interactions, but these systems still struggle with complex reasoning tasks. Previously, chain-of-thought (CoT) prompting or fine-tuning has been to shown to significantly improve the reasoning abilities of text-based LLMs. In this work, we investigate the effect of CoT fine-tuning for multi-stream speech LLMs, demonstrating that reasoning in text space improves the accuracy of speech LLMs by 2.4x, on average, over a suite of spoken reasoning tasks. Beyond accuracy, the latency of the spoken response is a crucial factor for interacting with voice-based agents. Inspired by the human behavior of "thinking while listening," we propose methods to reduce the additional latency from reasoning by allowing the model to start reasoning before the user query has ended. To achieve this, we introduce an entropy-based metric, "question completeness," which acts as an indicator to guide the model on the optimal time to start reasoning. This method provides greater control over the accuracy-latency trade-off compared with heuristic-based approaches and, under equivalent latency conditions, yields a 4% accuracy gain on ARC-Easy. Finally, we use Direct Preference Optimization (DPO) on preference data created using rejection sampling to push the accuracy-latency pareto frontier further, resulting in a 70% reduction in latency without loss in accuracy.

Yingyi Ma, Zhe Liu, Ozlem Kalinli

The advent of Large Language Models (LLM) has reformed the Automatic Speech Recognition (ASR). Prompting LLM with audio embeddings to generate transcriptions becomes the new state-of-the-art ASR. Despite LLMs being trained with an extensive amount of text corpora, high-quality domain-specific text data can still significantly enhance ASR performance on domain adaptation tasks. Although LLM-based ASR can naturally incorporate more text corpora by fine-tuning the LLM decoder, fine-tuning such ASR on text-only data without paired prompts may diminish the effectiveness of domain-specific knowledge. To mitigate this issue, we propose a two-step soft prompt fine-tuning strategy that enhances domain-specific text adaptation. Experimental results show that text adaptation with our proposed method achieved a relative up to 9% Word Error Rate (WER) reduction and up to 18% Entity Error Rate (EER) reduction on the target domain compared to the baseline ASR. Combining this with domain-specific Language Model (LM) fusion can further improve the EER by a relative 2-5%

Gil Keren, Wei Zhou, Ozlem Kalinli

ASR models are commonly trained with the cross-entropy criterion to increase the probability of a target token sequence. While optimizing the probability of all tokens in the target sequence is sensible, one may want to de-emphasize tokens that reflect transcription errors. In this work, we propose a novel token-weighted RNN-T criterion that augments the RNN-T objective with token-specific weights. The new objective is used for mitigating accuracy loss from transcriptions errors in the training data, which naturally appear in two settings: pseudo-labeling and human annotation errors. Experiments results show that using our method for semi-supervised learning with pseudo-labels leads to a consistent accuracy improvement, up to 38% relative. We also analyze the accuracy degradation resulting from different levels of WER in the reference transcription, and show that token-weighted RNN-T is suitable for overcoming this degradation, recovering 64%-99% of the accuracy loss.

Chuanneng Sun, Zeeshan Ahmed, Yingyi Ma et al.

This paper studies contextual biasing with Large Language Models (LLMs), where during second-pass rescoring additional contextual information is provided to a LLM to boost Automatic Speech Recognition (ASR) performance. We propose to leverage prompts for a LLM without fine tuning during rescoring which incorporate a biasing list and few-shot examples to serve as additional information when calculating the score for the hypothesis. In addition to few-shot prompt learning, we propose multi-task training of the LLM to predict both the entity class and the next token. To improve the efficiency for contextual biasing and to avoid exceeding LLMs' maximum sequence lengths, we propose dynamic prompting, where we select the most likely class using the class tag prediction, and only use entities in this class as contexts for next token prediction. Word Error Rate (WER) evaluation is performed on i) an internal calling, messaging, and dictation dataset, and ii) the SLUE-Voxpopuli dataset. Results indicate that biasing lists and few-shot examples can achieve 17.8% and 9.6% relative improvement compared to first pass ASR, and that multi-task training and dynamic prompting can achieve 20.0% and 11.3% relative WER improvement, respectively.

Shuo Liu, Leda Sarı, Chunyang Wu et al.

This paper presents a method for selecting appropriate synthetic speech samples from a given large text-to-speech (TTS) dataset as supplementary training data for an automatic speech recognition (ASR) model. We trained a neural network, which can be optimised using cross-entropy loss or Arcface loss, to measure the similarity of a synthetic data to real speech. We found that incorporating synthetic samples with considerable dissimilarity to real speech, owing in part to lexical differences, into ASR training is crucial for boosting recognition performance. Experimental results on Librispeech test sets indicate that, in order to maintain the same speech recognition accuracy as when using all TTS data, our proposed solution can reduce the size of the TTS data down below its $30\,\%$, which is superior to several baseline methods.

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.

Junteng Jia, Jay Mahadeokar, Weiyi Zheng et al.

Cross-device federated learning (FL) protects user privacy by collaboratively training a model on user devices, therefore eliminating the need for collecting, storing, and manually labeling user data. While important topics such as the FL training algorithm, non-IID-ness, and Differential Privacy have been well studied in the literature, this paper focuses on two challenges of practical importance for improving on-device ASR: the lack of ground-truth transcriptions and the scarcity of compute resource and network bandwidth on edge devices. First, we propose a FL system for on-device ASR domain adaptation with full self-supervision, which uses self-labeling together with data augmentation and filtering techniques. The system can improve a strong Emformer-Transducer based ASR model pretrained on out-of-domain data, using in-domain audio without any ground-truth transcriptions. Second, to reduce the training cost, we propose a self-restricted RNN Transducer (SR-RNN-T) loss, a variant of alignment-restricted RNN-T that uses Viterbi alignments from self-supervision. To further reduce the compute and network cost, we systematically explore adapting only a subset of weights in the Emformer-Transducer. Our best training recipe achieves a $12.9\%$ relative WER reduction over the strong out-of-domain baseline, which equals $70\%$ of the reduction achievable with full human supervision and centralized training.

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.

Antoine Bruguier, Duc Le, Rohit Prabhavalkar et al.

We propose Neural-FST Class Language Model (NFCLM) for end-to-end speech recognition, a novel method that combines neural network language models (NNLMs) and finite state transducers (FSTs) in a mathematically consistent framework. Our method utilizes a background NNLM which models generic background text together with a collection of domain-specific entities modeled as individual FSTs. Each output token is generated by a mixture of these components; the mixture weights are estimated with a separately trained neural decider. We show that NFCLM significantly outperforms NNLM by 15.8% relative in terms of Word Error Rate. NFCLM achieves similar performance as traditional NNLM and FST shallow fusion while being less prone to overbiasing and 12 times more compact, making it more suitable for on-device usage.

Alex Xiao, Weiyi Zheng, Gil Keren et al.

With 4.5 million hours of English speech from 10 different sources across 120 countries and models of up to 10 billion parameters, we explore the frontiers of scale for automatic speech recognition. We propose data selection techniques to efficiently scale training data to find the most valuable samples in massive datasets. To efficiently scale model sizes, we leverage various optimizations such as sparse transducer loss and model sharding. By training 1-10B parameter universal English ASR models, we push the limits of speech recognition performance across many domains. Furthermore, our models learn powerful speech representations with zero and few-shot capabilities on novel domains and styles of speech, exceeding previous results across multiple in-house and public benchmarks. For speakers with disorders due to brain damage, our best zero-shot and few-shot models achieve 22% and 60% relative improvement on the AphasiaBank test set, respectively, while realizing the best performance on public social media videos. Furthermore, the same universal model reaches equivalent performance with 500x less in-domain data on the SPGISpeech financial-domain dataset.

Haichuan Yang, Yuan Shangguan, Dilin Wang et al.

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.

Suyoun Kim, Duc Le, Weiyi Zheng et al.

Measuring automatic speech recognition (ASR) system quality is critical for creating user-satisfying voice-driven applications. Word Error Rate (WER) has been traditionally used to evaluate ASR system quality; however, it sometimes correlates poorly with user perception/judgement of transcription quality. This is because WER weighs every word equally and does not consider semantic correctness which has a higher impact on user perception. In this work, we propose evaluating ASR output hypotheses quality with SemDist that can measure semantic correctness by using the distance between the semantic vectors of the reference and hypothesis extracted from a pre-trained language model. Our experimental results of 71K and 36K user annotated ASR output quality show that SemDist achieves higher correlation with user perception than WER. We also show that SemDist has higher correlation with downstream Natural Language Understanding (NLU) tasks than WER.

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.

Dilin Wang, Yuan Shangguan, Haichuan Yang et al.

Automatic speech recognition (ASR) has become increasingly ubiquitous on modern edge devices. Past work developed streaming End-to-End (E2E) all-neural speech recognizers that can run compactly on edge devices. However, E2E ASR models are prone to overfitting and have difficulties in generalizing to unseen testing data. Various techniques have been proposed to regularize the training of ASR models, including layer normalization, dropout, spectrum data augmentation and speed distortions in the inputs. In this work, we present a simple yet effective noisy training strategy to further improve the E2E ASR model training. By introducing random noise to the parameter space during training, our method can produce smoother models at convergence that generalize better. We apply noisy training to improve both dense and sparse state-of-the-art Emformer models and observe consistent WER reduction. Specifically, when training Emformers with 90% sparsity, we achieve 12% and 14% WER improvements on the LibriSpeech Test-other and Test-clean data set, respectively.

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.

Suyoun Kim, Abhinav Arora, Duc Le et al.

Word Error Rate (WER) has been the predominant metric used to evaluate the performance of automatic speech recognition (ASR) systems. However, WER is sometimes not a good indicator for downstream Natural Language Understanding (NLU) tasks, such as intent recognition, slot filling, and semantic parsing in task-oriented dialog systems. This is because WER takes into consideration only literal correctness instead of semantic correctness, the latter of which is typically more important for these downstream tasks. In this study, we propose a novel Semantic Distance (SemDist) measure as an alternative evaluation metric for ASR systems to address this issue. We define SemDist as the distance between a reference and hypothesis pair in a sentence-level embedding space. To represent the reference and hypothesis as a sentence embedding, we exploit RoBERTa, a state-of-the-art pre-trained deep contextualized language model based on the transformer architecture. We demonstrate the effectiveness of our proposed metric on various downstream tasks, including intent recognition, semantic parsing, and named entity recognition.

Ting-Yao Hu, Ashish Shrivastava, Jen-Hao Rick Chang et al.

Data augmentation methods usually apply the same augmentation (or a mix of them) to all the training samples. For example, to perturb data with noise, the noise is sampled from a Normal distribution with a fixed standard deviation, for all samples. We hypothesize that a hard sample with high training loss already provides strong training signal to update the model parameters and should be perturbed with mild or no augmentation. Perturbing a hard sample with a strong augmentation may also make it too hard to learn from. Furthermore, a sample with low training loss should be perturbed by a stronger augmentation to provide more robustness to a variety of conditions. To formalize these intuitions, we propose a novel method to learn a Sample-Adaptive Policy for Augmentation -- SapAugment. Our policy adapts the augmentation parameters based on the training loss of the data samples. In the example of Gaussian noise, a hard sample will be perturbed with a low variance noise and an easy sample with a high variance noise. Furthermore, the proposed method combines multiple augmentation methods into a methodical policy learning framework and obviates hand-crafting augmentation parameters by trial-and-error. We apply our method on an automatic speech recognition (ASR) task, and combine existing and novel augmentations using the proposed framework. We show substantial improvement, up to 21% relative reduction in word error rate on LibriSpeech dataset, over the state-of-the-art speech augmentation method.

Gautam Mantena, Ozlem Kalinli, Ossama Abdel-Hamid et al.

In this paper, we tackle the problem of handling narrowband and wideband speech by building a single acoustic model (AM), also called mixed bandwidth AM. In the proposed approach, an auxiliary input feature is used to provide the bandwidth information to the model, and bandwidth embeddings are jointly learned as part of acoustic model training. Experimental evaluations show that using bandwidth embeddings helps the model to handle the variability of the narrow and wideband speech, and makes it possible to train a mixed-bandwidth AM. Furthermore, we propose to use parallel convolutional layers to handle the mismatch between the narrow and wideband speech better, where separate convolution layers are used for each type of input speech signal. Our best system achieves 13% relative improvement on narrowband speech, while not degrading on wideband speech.