Contextual biasing is an important and challenging task for end-to-end automatic speech recognition (ASR) systems, which aims to achieve better recognition performance by biasing the ASR system to particular context phrases such as person names, music list, proper nouns, etc. Existing methods mainly include contextual LM biasing and adding bias encoder into end-to-end ASR models. In this work, we introduce a novel approach to do contextual biasing by adding a contextual spelling correction model on top of the end-to-end ASR system. We incorporate contextual information into a sequence-to-sequence spelling correction model with a shared context encoder. Our proposed model includes two different mechanisms: autoregressive (AR) and non-autoregressive (NAR). We propose filtering algorithms to handle large-size context lists, and performance balancing mechanisms to control the biasing degree of the model. We demonstrate the proposed model is a general biasing solution which is domain-insensitive and can be adopted in different scenarios. Experiments show that the proposed method achieves as much as 51% relative word error rate (WER) reduction over ASR system and outperforms traditional biasing methods. Compared to the AR solution, the proposed NAR model reduces model size by 43.2% and speeds up inference by 2.1 times.
The Transformer architecture has been well adopted as a dominant architecture in most sequence transduction tasks including automatic speech recognition (ASR), since its attention mechanism excels in capturing long-range dependencies. While models built solely upon attention can be better parallelized than regular RNN, a novel network architecture, SRU++, was recently proposed. By combining the fast recurrence and attention mechanism, SRU++ exhibits strong capability in sequence modeling and achieves near-state-of-the-art results in various language modeling and machine translation tasks with improved compute efficiency. In this work, we present the advantages of applying SRU++ in ASR tasks by comparing with Conformer across multiple ASR benchmarks and study how the benefits can be generalized to long-form speech inputs. On the popular LibriSpeech benchmark, our SRU++ model achieves 2.0% / 4.7% WER on test-clean / test-other, showing competitive performances compared with the state-of-the-art Conformer encoder under the same set-up. Specifically, SRU++ can surpass Conformer on long-form speech input with a large margin, based on our analysis.
Self-attention (SA), which encodes vector sequences according to their pairwise similarity, is widely used in speech recognition due to its strong context modeling ability. However, when applied to long sequence data, its accuracy is reduced. This is caused by the fact that its weighted average operator may lead to the dispersion of the attention distribution, which results in the relationship between adjacent signals ignored. To address this issue, in this paper, we introduce relative-position-awareness self-attention (RPSA). It not only maintains the global-range dependency modeling ability of self-attention, but also improves the localness modeling ability. Because the local window length of the original RPSA is fixed and sensitive to different test data, here we propose Gaussian-based self-attention (GSA) whose window length is learnable and adaptive to the test data automatically. We further generalize GSA to a new residual Gaussian self-attention (resGSA) for the performance improvement. We apply RPSA, GSA, and resGSA to Transformer-based speech recognition respectively. Experimental results on the AISHELL-1 Mandarin speech recognition corpus demonstrate the effectiveness of the proposed methods. For example, the resGSA-Transformer achieves a character error rate (CER) of 5.86% on the test set, which is relative 7.8% lower than that of the SA-Transformer. Although the performance of the proposed resGSA-Transformer is only slightly better than that of the RPSA-Transformer, it does not have to tune the window length manually.
Inspired by the progress of the End-to-End approach [1], this paper systematically studies the effects of Number of Filters of convolutional layers on the model prediction accuracy of CNN+RNN (Convolutional Neural Networks adding to Recurrent Neural Networks) for ASR Models (Automatic Speech Recognition). Experimental results show that only when the CNN Number of Filters exceeds a certain threshold value is adding CNN to RNN able to improve the performance of the CNN+RNN speech recognition model, otherwise some parameter ranges of CNN can render it useless to add the CNN to the RNN model. Our results show a strong dependency of word accuracy on the Number of Filters of convolutional layers. Based on the experimental results, the paper suggests a possible hypothesis of Sound-2-Vector Embedding (Convolutional Embedding) to explain the above observations. Based on this Embedding hypothesis and the optimization of parameters, the paper develops an End-to-End speech recognition system which has a high word accuracy but also has a light model-weight. The developed LVCSR (Large Vocabulary Continuous Speech Recognition) model has achieved quite a high word accuracy of 90.2% only by its Acoustic Model alone, without any assistance from intermediate phonetic representation and any Language Model. Its acoustic model contains only 4.4 million weight parameters, compared to the 35~68 million acoustic-model weight parameters in DeepSpeech2 [2] (one of the top state-of-the-art LVCSR models) which can achieve a word accuracy of 91.5%. The light-weighted model is good for improving the transcribing computing efficiency and also useful for mobile devices, Driverless Vehicles, etc. Our model weight is reduced to ~10% the size of DeepSpeech2, but our model accuracy remains close to that of DeepSpeech2. If combined with a Language Model, our LVCSR system is able to achieve 91.5% word accuracy.
Recently, attention-based encoder-decoder (AED) models have shown high performance for end-to-end automatic speech recognition (ASR) across several tasks. Addressing overconfidence in such models, in this paper we introduce the concept of relaxed attention, which is a simple gradual injection of a uniform distribution to the encoder-decoder attention weights during training that is easily implemented with two lines of code. We investigate the effect of relaxed attention across different AED model architectures and two prominent ASR tasks, Wall Street Journal (WSJ) and Librispeech. We found that transformers trained with relaxed attention outperform the standard baseline models consistently during decoding with external language models. On WSJ, we set a new benchmark for transformer-based end-to-end speech recognition with a word error rate of 3.65%, outperforming state of the art (4.20%) by 13.1% relative, while introducing only a single hyperparameter. Upon acceptance, models will be published on github.
In the recent years end to end (E2E) automatic speech recognition (ASR) systems have achieved promising results given sufficient resources. Even for languages where not a lot of labelled data is available, state of the art E2E ASR systems can be developed by pretraining on huge amounts of high resource languages and finetune on low resource languages. For a lot of low resource languages the current approaches are still challenging, since in many cases labelled data is not available in open domain. In this paper we present an approach to create labelled data for Maithili, Bhojpuri and Dogri by utilising pseudo labels from text to speech for forced alignment. The created data was inspected for quality and then further used to train a transformer based wav2vec 2.0 ASR model. All data and models are available in open domain.
Code-switching in automatic speech recognition (ASR) is an important challenge due to globalization. Recent research in multilingual ASR shows potential improvement over monolingual systems. We study key issues related to multilingual modeling for ASR through a series of large-scale ASR experiments. Our innovative framework deploys a multi-graph approach in the weighted finite state transducers (WFST) framework. We compare our WFST decoding strategies with a transformer sequence to sequence system trained on the same data. Given a code-switching scenario between Arabic and English languages, our results show that the WFST decoding approaches were more suitable for the intersentential code-switching datasets. In addition, the transformer system performed better for intrasentential code-switching task. With this study, we release an artificially generated development and test sets, along with ecological code-switching test set, to benchmark the ASR performance.
Personal voice activity detection has received increased attention due to the growing popularity of personal mobile devices and smart speakers. PVAD is often an integral element to speech enhancement and recognition for these applications in which lightweight signal processing is only enabled for the target user. However, in real-world scenarios, the detection performance may degrade because of competing speakers, background noise, and reverberation. To address this problem, we proposed to use equivalent rectangular bandwidth ERB-scaled spatial coherence as the input feature to train an array configuration-agnostic PVAD network. Whereas the network model requires only 112k parameters, it exhibits excellent detection performance and robustness in adverse acoustic conditions. Notably, the proposed ARCA-PVAD system is scalable to array configurations. Experimental results have demonstrated the superior performance achieved by the proposed ARCA-PVAD system over a baseline in terms of the area under receiver operating characteristic curve and equal error rate.
Machine learning methods for conditional data generation usually build a mapping from source conditional data X to target data Y. The target Y (e.g., text, speech, music, image, video) is usually high-dimensional and complex, and contains information that does not exist in source data, which hinders effective and efficient learning on the source-target mapping. In this paper, we present a learning paradigm called regeneration learning for data generation, which first generates Y' (an abstraction/representation of Y) from X and then generates Y from Y'. During training, Y' is obtained from Y through either handcrafted rules or self-supervised learning and is used to learn X-->Y' and Y'-->Y. Regeneration learning extends the concept of representation learning to data generation tasks, and can be regarded as a counterpart of traditional representation learning, since 1) regeneration learning handles the abstraction (Y') of the target data Y for data generation while traditional representation learning handles the abstraction (X') of source data X for data understanding; 2) both the processes of Y'-->Y in regeneration learning and X-->X' in representation learning can be learned in a self-supervised way (e.g., pre-training); 3) both the mappings from X to Y' in regeneration learning and from X' to Y in representation learning are simpler than the direct mapping from X to Y. We show that regeneration learning can be a widely-used paradigm for data generation (e.g., text generation, speech recognition, speech synthesis, music composition, image generation, and video generation) and can provide valuable insights into developing data generation methods.