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

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

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

The traditional hybrid deep neural network (DNN)–hidden Markov model (HMM) system and attention-based encoder–decoder (AED) model are both commonly used automatic speech recognition (ASR) approaches with distinct characteristics and advantages. While hybrid systems are per-frame-based and highly modularised to leverage external phonetic and linguistic knowledge, AED models operate on a per-label basis and jointly learn the acoustic and language information using a single model in an end-to-end trainable fashion. In this paper, we propose combining these two approaches in a two-pass rescoring framework. The first-pass uses hybrid ASR systems to facilitate streaming and controllable ASR, and the second-pass re-scores the N-best hypotheses or lattices produced by the first-pass hybrid DNN-HMM system with AED models. We also propose an improved algorithm for lattice rescoring with AED models. Experiments show the combined two-pass systems achieve competitive performance without using extra speech or text data on two standard ASR tasks. For the 80-hour AMI IHM dataset, the combined system has a 13.7% word error rate (WER) on the evaluation set and is up to a 29% relative WER reduction over the individual systems. For the 300-hour Switchboard dataset, the WERs of the combined system are 5.7% and 12.1% on Switchboard and CallHome subsets of Hub5’00, and 13.2% and 7.6% on Switchboard Cellular and Fisher subsets of RT03, and are up to a 33% relative reduction in WER over the individual systems.

As end-to-end automatic speech recognition (ASR) models reach promising performance, various downstream tasks rely on good confidence estimators for these systems. Recent research has shown that model-based confidence estimators have a significant advantage over using the output softmax probabilities. If the input data to the speech recogniser is from mismatched acoustic and linguistic conditions, the ASR performance and the corresponding confidence estimators may exhibit severe degradation. Since confidence models are often trained on the same in-domain data as the ASR, generalising to out-of-domain (OOD) scenarios is challenging. By keeping the ASR model untouched, this paper proposes two approaches to improve the model-based confidence estimators on OOD data: using pseudo transcriptions and an additional OOD language model. With an ASR model trained on LibriSpeech, experiments show that the proposed methods can greatly improve the confidence metrics on TED-LIUM and Switchboard datasets while preserving in-domain performance. Furthermore, the improved confidence estimators are better calibrated on OOD data and can provide a much more reliable criterion for data selection.

Self-supervised pre-training is an effective approach to leveraging a large amount of unlabelled data to reduce word error rates (WERs) of automatic speech recognition (ASR) systems. Since it is impractical to use large pre-trained models for many real-world ASR applications, it is desirable to have a much smaller model while retaining the performance of the pre-trained model. In this paper, we propose a simple knowledge distillation (KD) loss function for neural transducers that focuses on the one-best path in the output probability lattice under both streaming and non-streaming setups, which allows a small student model to approach the performance of the large pre-trained teacher model. Experiments on the LibriSpeech dataset show that despite being 10 times smaller than the teacher model, the proposed loss results in relative WER reductions (WERRs) of 11.5% and 6.8% on the test-other set for non-streaming and streaming student models compared to the baseline transducers trained without KD using the labelled 100-hour clean data. With an additional 860 hours of unlabelled data for KD, the WERRs increase to 48.2% and 38.5% for non-streaming and streaming students. If language model shallow fusion is used for producing distillation targets, a further improvement in the student model is observed.

Accurate confidence measures for predictions from machine learning techniques play a critical role in the deployment and training of many speech and language processing applications. For example, confidence scores are important when making use of automatically generated transcriptions in training automatic speech recognition (ASR) systems, as well as down-stream applications, such as information retrieval and conversational assistants. Previous work on improving confidence scores for these systems has focused on two main directions: designing features correlated with improved confidence prediction; and employing sequence models to account for the importance of contextual information. Few studies, however, have explored incorporating contextual information more broadly, such as from the future, in addition to the past, or making use of alternative multiple hypotheses in addition to the most likely one. This article introduces two general approaches for encapsulating contextual information from lattices. Experimental results illustrating the importance of increasing contextual information for estimating confidence scores are presented on a range of limited resource languages where word error rates range between 30% and 60%. The results show that the novel approaches provide significant gains in the accuracy of confidence estimation.

For various speech-related tasks, confidence scores from a speech recogniser are a useful measure to assess the quality of transcriptions. In traditional hidden Markov model-based automatic speech recognition (ASR) systems, confidence scores can be reliably obtained from word posteriors in decoding lattices. However, for an ASR system with an auto-regressive decoder, such as an attention-based sequence-to-sequence model, computing word posteriors is difficult. An obvious alternative is to use the decoder softmax probability as the model confidence. In this paper, we first examine how some commonly used regularisation methods influence the softmax-based confidence scores and study the overconfident behaviour of end-to-end models. Then we propose a lightweight and effective approach named confidence estimation module (CEM) on top of an existing end-to-end ASR model. Experiments on LibriSpeech show that CEM can mitigate the overconfidence problem and can produce more reliable confidence scores with and without shallow fusion of a language model. Further analysis shows that CEM generalises well to speech from a moderately mismatched domain and can potentially improve downstream tasks such as semi-supervised learning.

We study the problem of word-level confidence estimation in subword-based end-to-end (E2E) models for automatic speech recognition (ASR). Although prior works have proposed training auxiliary confidence models for ASR systems, they do not extend naturally to systems that operate on word-pieces (WP) as their vocabulary. In particular, ground truth WP correctness labels are needed for training confidence models, but the non-unique tokenization from word to WP causes inaccurate labels to be generated. This paper proposes and studies two confidence models of increasing complexity to solve this problem. The final model uses self-attention to directly learn word-level confidence without needing subword tokenization, and exploits full context features from multiple hypotheses to improve confidence accuracy. Experiments on Voice Search and long-tail test sets show standard metrics (e.g., NCE, AUC, RMSE) improving substantially. The proposed confidence module also enables a model selection approach to combine an on-device E2E model with a hybrid model on the server to address the rare word recognition problem for the E2E model.

End-to-end models with auto-regressive decoders have shown impressive results for automatic speech recognition (ASR). These models formulate the sequence-level probability as a product of the conditional probabilities of all individual tokens given their histories. However, the performance of locally normalised models can be sub-optimal because of factors such as exposure bias. Consequently, the model distribution differs from the underlying data distribution. In this paper, the residual energy-based model (R-EBM) is proposed to complement the auto-regressive ASR model to close the gap between the two distributions. Meanwhile, R-EBMs can also be regarded as utterance-level confidence estimators, which may benefit many downstream tasks. Experiments on a 100hr LibriSpeech dataset show that R-EBMs can reduce the word error rates (WERs) by 8.2%/6.7% while improving areas under precision-recall curves of confidence scores by 12.6%/28.4% on test-clean/test-other sets. Furthermore, on a state-of-the-art model using self-supervised learning (wav2vec 2.0), R-EBMs still significantly improves both the WER and confidence estimation performance.

Confidence scores are very useful for downstream applications of automatic speech recognition (ASR) systems. Recent works have proposed using neural networks to learn word or utterance confidence scores for end-to-end ASR. In those studies, word confidence by itself does not model deletions, and utterance confidence does not take advantage of word-level training signals. This paper proposes to jointly learn word confidence, word deletion, and utterance confidence. Empirical results show that multi-task learning with all three objectives improves confidence metrics (NCE, AUC, RMSE) without the need for increasing the model size of the confidence estimation module. Using the utterance-level confidence for rescoring also decreases the word error rates on Google's Voice Search and Long-tail Maps datasets by 3-5% relative, without needing a dedicated neural rescorer.

In this paper, we propose Discriminative Neural Clustering (DNC) that formulates data clustering with a maximum number of clusters as a supervised sequence-to-sequence learning problem. Com-pared to traditional unsupervised clustering algorithms, DNC learns clustering patterns from training data without requiring an explicit definition of a similarity measure. An implementation of DNC based on the Transformer architecture is shown to be effective on a speaker diarisation task using the challenging AMI dataset. Since AMI contains only 147 complete meetings as individual input sequences, data scarcity is a significant issue for training a Transformer model for DNC. Accordingly, this paper proposes three data augmentation schemes: sub-sequence randomisation, input vector randomisation, and Diaconis augmentation, which generates new data samples by rotating the entire input sequence of L2-normalised speaker embeddings. Experimental results on AMI show that DNC achieves a reduction in speaker error rate (SER) of 29.4% relative to spectral clustering.

This paper proposes a novel automatic speech recognition (ASR) framework called Integrated Source-Channel and Attention (ISCA) that combines the advantages of traditional systems based on the noisy source-channel model (SC) and end-to-end style systems using attention-based sequence-to-sequence models. The traditional SC system framework includes hidden Markov models and connectionist temporal classification (CTC) based acoustic models, language models (LMs), and a decoding procedure based on a lexicon, whereas the end-to-end style attention-based system jointly models the whole process with a single model. By rescoring the hypotheses produced by traditional systems using end-to-end style systems based on an extended noisy source-channel model, ISCA allows structured knowledge to be easily incorporated via the SC-based model while exploiting the complementarity of the attention-based model. Experiments on the AMI meeting corpus show that ISCA is able to give a relative word error rate reduction up to 21% over an individual system, and by 13% over an alternative method which also involves combining CTC and attention-based models.

The standard approach to mitigate errors made by an automatic speech recognition system is to use confidence scores associated with each predicted word. In the simplest case, these scores are word posterior probabilities whilst more complex schemes utilise bi-directional recurrent neural network (BiRNN) models. A number of upstream and downstream applications, however, rely on confidence scores assigned not only to 1-best hypotheses but to all words found in confusion networks or lattices. These include but are not limited to speaker adaptation, semi-supervised training and information retrieval. Although word posteriors could be used in those applications as confidence scores, they are known to have reliability issues. To make improved confidence scores more generally available, this paper shows how BiRNNs can be extended from 1-best sequences to confusion network and lattice structures. Experiments are conducted using one of the Cambridge University submissions to the IARPA OpenKWS 2016 competition. The results show that confusion network and lattice-based BiRNNs can provide a significant improvement in confidence estimation.

This paper describes PyHTK, which is a Python-based library and associated pipeline to facilitate the construction of large-scale complex automatic speech recognition (ASR) systems using the hidden Markov model toolkit (HTK). PyHTK can be used to generate sophisticated artificial neural network (ANN) models with versatile architectures by converting a compact configuration file defining the ANN, into the form used by HTK tools, as well as supporting a range of capabilities to train and test ANN models. The ASR pipeline is divided into multiple steps, which can be arranged and customised for different ASR data sets, and allows for both step-by-step and fully automatic end-to-end operation. PyHTK is integrated with HTK 3.5.1 which includes an expanded range of ANN layer types and very flexible ways to connect them, together with capabilities for ASR training and testing. Some example systems are included to illustrate the flexibility and performance achievable.

The standard approach to assess reliability of automatic speech transcriptions is through the use of confidence scores. If accurate, these scores provide a flexible mechanism to flag transcription errors for upstream and downstream applications. One challenging type of errors that recognisers make are deletions. These errors are not accounted for by the standard confidence estimation schemes and are hard to rectify in the upstream and downstream processing. High deletion rates are prominent in limited resource and highly mismatched training/testing conditions studied under IARPA Babel and Material programs. This paper looks at the use of bidirectional recurrent neural networks to yield confidence estimates in predicted as well as deleted words. Several simple schemes are examined for combination. To assess usefulness of this approach, the combined confidence score is examined for untranscribed data selection that favours transcriptions with lower deletion errors. Experiments are conducted using IARPA Babel/Material program languages.

Humans infer rich knowledge of objects from both auditory and visual cues. Building a machine of such competency, however, is very challenging, due to the great difficulty in capturing large-scale, clean data of objects with both their appearance and the sound they make. In this paper, we present a novel, open-source pipeline that generates audiovisual data, purely from 3D object shapes and their physical properties. Through comparison with audio recordings and human behavioral studies, we validate the accuracy of the sounds it generates. Using this generative model, we are able to construct a synthetic audio-visual dataset, namely Sound-20K, for object perception tasks. We demonstrate that auditory and visual information play complementary roles in object perception, and further, that the representation learned on synthetic audio-visual data can transfer to real-world scenarios.

Hearing an object falling onto the ground, humans can recover rich information including its rough shape, material, and falling height. In this paper, we build machines to approximate such competency. We first mimic human knowledge of the physical world by building an efficient, physics-based simulation engine. Then, we present an analysis-by-synthesis approach to infer properties of the falling object. We further accelerate the process by learning a mapping from a sound wave to object properties, and using the predicted values to initialize the inference. This mapping can be viewed as an approximation of human commonsense learned from past experience. Our model performs well on both synthetic audio clips and real recordings without requiring any annotated data. We conduct behavior studies to compare human responses with ours on estimating object shape, material, and falling height from sound. Our model achieves near-human performance.