End-to-end speech recognition is a promising technology for enabling compact automatic speech recognition (ASR) systems since it can unify the acoustic and language model into a single neural network. However, as a drawback, training of end-to-end speech recognizers always requires transcribed utterances. Since end-to-end models are also known to be severely data hungry, this constraint is crucial especially because obtaining transcribed utterances is costly and can possibly be impractical or impossible. This paper proposes a method for alleviating this issue by transferring knowledge from a language model neural network that can be pretrained with text-only data. Specifically, this paper attempts to transfer semantic knowledge acquired in embedding vectors of large-scale language models. Since embedding vectors can be assumed as implicit representations of linguistic information such as part-of-speech, intent, and so on, those are also expected to be useful modeling cues for ASR decoders. This paper extends two types of ASR decoders, attention-based decoders and neural transducers, by modifying training loss functions to include embedding prediction terms. The proposed systems were shown to be effective for error rate reduction without incurring extra computational costs in the decoding phase.

End-to-end (E2E) modeling is advantageous for automatic speech recognition (ASR) especially for Japanese since word-based tokenization of Japanese is not trivial, and E2E modeling is able to model character sequences directly. This paper focuses on the latest E2E modeling techniques, and investigates their performances on character-based Japanese ASR by conducting comparative experiments. The results are analyzed and discussed in order to understand the relative advantages of long short-term memory (LSTM), and Conformer models in combination with connectionist temporal classification, transducer, and attention-based loss functions. Furthermore, the paper investigates on effectivity of the recent training techniques such as data augmentation (SpecAugment), variational noise injection, and exponential moving average. The best configuration found in the paper achieved the state-of-the-art character error rates of 4.1%, 3.2%, and 3.5% for Corpus of Spontaneous Japanese (CSJ) eval1, eval2, and eval3 tasks, respectively. The system is also shown to be computationally efficient thanks to the efficiency of Conformer transducers.

We present an approach for unsupervised learning of speech representation disentangling contents and styles. Our model consists of: (1) a local encoder that captures per-frame information; (2) a global encoder that captures per-utterance information; and (3) a conditional decoder that reconstructs speech given local and global latent variables. Our experiments show that (1) the local latent variables encode speech contents, as reconstructed speech can be recognized by ASR with low word error rates (WER), even with a different global encoding; (2) the global latent variables encode speaker style, as reconstructed speech shares speaker identity with the source utterance of the global encoding. Additionally, we demonstrate an useful application from our pre-trained model, where we can train a speaker recognition model from the global latent variables and achieve high accuracy by fine-tuning with as few data as one label per speaker.

Sequence-to-sequence models have been widely used in end-to-end speech processing, for example, automatic speech recognition (ASR), speech translation (ST), and text-to-speech (TTS). This paper focuses on an emergent sequence-to-sequence model called Transformer, which achieves state-of-the-art performance in neural machine translation and other natural language processing applications. We undertook intensive studies in which we experimentally compared and analyzed Transformer and conventional recurrent neural networks (RNN) in a total of 15 ASR, one multilingual ASR, one ST, and two TTS benchmarks. Our experiments revealed various training tips and significant performance benefits obtained with Transformer for each task including the surprising superiority of Transformer in 13/15 ASR benchmarks in comparison with RNN. We are preparing to release Kaldi-style reproducible recipes using open source and publicly available datasets for all the ASR, ST, and TTS tasks for the community to succeed our exciting outcomes.

We introduce speech and text autoencoders that share encoders and decoders with an automatic speech recognition (ASR) model to improve ASR performance with large speech only and text only training datasets. To build the speech and text autoencoders, we leverage state-of-the-art ASR and text-to-speech (TTS) encoder decoder architectures. These autoencoders learn features from speech only and text only datasets by switching the encoders and decoders used in the ASR and TTS models. Simultaneously, they aim to encode features to be compatible with ASR and TTS models by a multi-task loss. Additionally, we anticipate that TTS joint training can also improve the ASR performance because both ASR and TTS models learn transformations between speech and text. The experimental result we obtained with our semi-supervised end-to-end ASR/TTS training revealed reductions from a model initially trained with a small paired subset of the LibriSpeech corpus in the character error rate from 10.4% to 8.4% and word error rate from 20.6% to 18.0% by retraining the model with a large unpaired subset of the corpus.

Beamforming approaches using time-frequency masks have recently been investigated and have shown promising results for noise robust automatic speech recognition (ASR) in many tasks. The time-frequency masks are estimated to compute the spatial statistics of target speech and noise signals, and then the statistics are used to derive a beamformer. Although its effectiveness has been clearly shown in batch and blockwise processing, it has not been well extended to frame-by-frame processing, which is a very important procedure for many actual applications. In this paper, we derive a frame-by-frame update rule for a mask-based minimum variance distortion-less response (MVDR) beamformer, which enables us to obtain enhanced signals without a long delay by combining it with uni-directional recurrent neural network-based mask estimation. Based on the Woodbury matrix identity, our algorithm achieves a closed-form solution of the mask-based MVDR beamformer at every time frame without any matrix inversion. Experimental results show that our frame-by-frame beamformer outperforms baseline block-wise beamforming on the CHiME-3 simulation dataset even with a shorter time delay.

The standard evaluation metric of automatic speech recognition (ASR) is the word error rate (WER), which measures the dissimilarity between recognized word sequences and their ground truth. Many training algorithms designed to reduce sequence-level errors such as WER have been proposed for hidden Markov model (HMM)-based ASR, e.g., state-level minimum Bayes risk (sMBR). However, these approaches cannot be used directly for encoder-decoder model based end-to-end ASR, because the encoder-decoder model employs very different mechanisms from HMM-based approaches. In this paper, we propose a new method for optimizing the encoder-decoder model based on a sequence-level evaluation metric. Since the WER is not directly differentiable, we adopt a policy gradient objective function to train the encoder-decoder model, which enables us to minimize the expected WER of the model predictions. This training method employs the scoring of multiple hypotheses as in the decoding stage while usual cross entropy training uses only the ground truth. Therefore, we can expect it to improve the decoding results of the encoder-decoder model. We perform experiments using the Tedlium corpus to demonstrate the potential of our proposed method for improving the recognition performance of the encoder-decoder model.

This paper proposes a new model for accurately rescoring (reranking) N-best speech recognition hypothesis lists. The model is based on state-of-the-art neural networks (NNs) and provides the minimum necessary functionality to perform N-best rescoring, i.e. one-on-one hypothesis comparison on a given N-best list in terms of word error rates (WERs). The model is composed of a long short-term memory (LSTM)-based encoder network followed by a fully-connected feedforward NN-based binary-class classifier network. Given the feature vector sequences of two hypotheses to be compared, this encoder-classifier (EC) model encodes these features and outputs binary-class probabilities that indicate which hypothesis has the lower WER. Then, depending on the output, the ranks of these hypotheses can be swapped. By repeating this one-on-one hypothesis comparison, a reranked N-best list can be obtained. In N-best rescoring experiments using a large scale speech corpus, the proposed EC model steadily outperforms an LSTM-based language model (LSTMLM), which is a strong and widely-used competitor. In addition, by incorporating the LSTMLM scores as an additional feature vector dimension, the N-best rescoring performance of the EC model is further improved. The improved EC model achieves a 10% relative WER reduction from the LSTMLM baseline.

This paper introduces a new open source platform for end-to-end speech processing named ESPnet. ESPnet mainly focuses on end-to-end automatic speech recognition (ASR), and adopts widely-used dynamic neural network toolkits, Chainer and PyTorch, as a main deep learning engine. ESPnet also follows the Kaldi ASR toolkit style for data processing, feature extraction/format, and recipes to provide a complete setup for speech recognition and other speech processing experiments. This paper explains a major architecture of this software platform, several important functionalities, which differentiate ESPnet from other open source ASR toolkits, and experimental results with major ASR benchmarks.

Recent studies in speech recognition have shown that the performance of convolutional neural networks (CNNs) is superior to that of fully connected deep neural networks (DNNs). In this paper, we explore the use of CNNs in far-field speech recognition for dealing with reverberation, which blurs spectral energies along the time axis. Unlike most previous CNN applications to speech recognition, we consider convolution in time to examine whether it provides an improved reverberation modelling capability. Experimental results show that a CNN coupled with a fully connected DNN can model short time correlations in feature vectors with fewer parameters than a DNN and thus generalise better to unseen test environments. Combining this approach with signal-space dereverberation, which copes with long-term correlations, is shown to result in further improvement, where the gains from both approaches are almost additive. An initial investigation of the use of restricted convolution forms is also undertaken.