Much of the recent progress in automatic speech recognition (ASR) lies in developing an acoustic encoder, such as enlarging its capacity and designing a refined architecture for speech processing. With these highly optimized encoders, the decoder has become less influential in its role as a language model (LM). In this work, we explore an effective approach for employing the LM structure in an ASR model. The proposed Mask-Conformer augments a Conformer-based model with a mask-predict decoder, which learns output context via the masked LM objective. The mask-predict decoder is applied to stacks of encoder layers, where the decoder output explicitly conditions the subsequent layers using cross-attention. We also propose a fill-mask decoding algorithm that refines a sequence using the decoder’s linguistic information. Experimental results show that Mask-Conformer outperforms strong baselines on some tasks. In addition, our analyses validate the effectiveness of the proposed model design.

This work studies knowledge distillation (KD) and addresses its constraints for recurrent neural network transducer (RNN-T) models. In hard distillation, a teacher model transcribes large amounts of unlabelled speech to train a student model. Soft distillation is another popular KD method that distills the output logits of the teacher model. Due to the nature of RNN-T alignments, applying soft distillation between RNNT architectures having different posterior distributions is challenging. In addition, bad teachers having high word-error-rate (WER) reduce the efficacy of KD. We investigate how to effectively distill knowledge from variable quality ASR teachers, which has not been studied before to the best of our knowledge. We show that a sequence-level KD, full-sum distillation, outperforms other distillation methods for RNN-T models, especially for bad teachers. We also propose a variant of full-sum distillation that distills the sequence discriminative knowledge of the teacher leading to further improvement in WER. We conduct experiments on public datasets namely SpeechStew and LibriSpeech, and on in-house production data.

Self-supervised ASR-TTS models suffer in out-of-domain data conditions. Here we propose an enhanced ASR-TTS (EAT) model that incorporates two main features: 1) The ASR→TTS direction is equipped with a language model reward to penalize the ASR hypotheses before forwarding it to TTS. 2) In the TTS→ASR direction, a hyper-parameter is introduced to scale the attention context from synthesized speech before sending it to ASR to handle out-of-domain data. Training strategies and the effectiveness of the EAT model are explored under out-of-domain data conditions. The results show that EAT reduces the performance gap between supervised and self-supervised training significantly by absolute 2.6% and 2.7% on Librispeech and BABEL respectively.

In this paper, we present promising accurate prefix boosting (PAPB), a discriminative training technique for attention based sequence-to-sequence (seq2seq) ASR. PAPB is devised to unify the training and testing scheme effectively. The training procedure involves maximizing the score of each partial correct sequence obtained during beam search compared to other hypotheses. The training objective also includes minimization of token (character) error rate. PAPB shows its efficacy by achieving 10.8% and 3.8% WER with and without external RNNLM respectively on Wall Street Journal dataset.

In this paper, we explore several new schemes to train a seq2seq model to integrate a pre-trained language model (LM). Our proposed fusion methods focus on the memory cell state and the hidden state in the seq2seq decoder long short-term memory (LSTM), and the memory cell state is updated by the LM unlike the prior studies. This means the memory retained by the main seq2seq would be adjusted by the external LM. These fusion methods have several variants depending on the architecture of this memory cell update and the use of memory cell and hidden states which directly affects the final label inference. We performed the experiments to show the effectiveness of the proposed methods in a mono-lingual ASR setup on the Librispeech corpus and in a transfer learning setup from a multilingual ASR (MLASR) base model to a low-resourced language. In Librispeech, our best model improved WER by 3.7%, 2.4% for test clean, test other relatively to the shallow fusion baseline, with multilevel decoding. In transfer learning from an MLASR base model to the IARPA Babel Swahili model, the best scheme improved the transferred model on eval set by 9.9%, 9.8% in CER, WER relatively to the 2-stage transfer baseline.

This work explores better adaptation methods to low-resource languages using an external language model (LM) under the framework of transfer learning. We first build a language-independent ASR system in a unified sequence-to-sequence (S2S) architecture with a shared vocabulary among all languages. During adaptation, we perform LM fusion transfer, where an external LM is integrated into the decoder network of the attention-based S2S model in the whole adaptation stage, to effectively incorporate linguistic context of the target language. We also investigate various seed models for transfer learning. Experimental evaluations using the IARPA BABEL data set show that LM fusion transfer improves performances on all target five languages compared with simple transfer learning when the external text data is available. Our final system drastically reduces the performance gap from the hybrid systems.

Sequence-to-sequence automatic speech recognition (ASR) models require large quantities of data to attain high performance. For this reason, there has been a recent surge in interest for unsupervised and semi-supervised training in such models. This work builds upon recent results showing notable improvements in semi-supervised training using cycle-consistency and related techniques. Such techniques derive training procedures and losses able to leverage unpaired speech and/or text data by combining ASR with Text-to-Speech (TTS) models. In particular, this work proposes a new semi-supervised loss combining an end-to-end differentiable ASR$\rightarrow$TTS loss with TTS$\rightarrow$ASR loss. The method is able to leverage both unpaired speech and text data to outperform recently proposed related techniques in terms of \%WER. We provide extensive results analyzing the impact of data quantity and speech and text modalities and show consistent gains across WSJ and Librispeech corpora. Our code is provided in ESPnet to reproduce the experiments.

This paper investigates the applications of various multilingual approaches developed in conventional hidden Markov model (HMM) systems to sequence-to-sequence (seq2seq) automatic speech recognition (ASR). On a set composed of Babel data, we first show the effectiveness of multi-lingual training with stacked bottle-neck (SBN) features. Then we explore various architectures and training strategies of multi-lingual seq2seq models based on CTC-attention networks including combinations of output layer, CTC and/or attention component re-training. We also investigate the effectiveness of language-transfer learning in a very low resource scenario when the target language is not included in the original multi-lingual training data. Interestingly, we found multilingual features superior to multilingual models, and this finding suggests that we can efficiently combine the benefits of the HMM system with the seq2seq system through these multilingual feature techniques.

The paper provides an analysis of automatic speech recognition systems (ASR) based on multilingual BLSTM, where we used multi-task training with separate classification layer for each language. The focus is on low resource languages, where only a limited amount of transcribed speech is available. In such scenario, we found it essential to train the ASR systems in a multilingual fashion and we report superior results obtained with pre-trained multilingual BLSTM on this task. The high resource languages are also taken into account and we show the importance of language richness for multilingual training. Next, we present the performance of this technique as a function of amount of target language data. The importance of including context information into BLSTM multilingual systems is also stressed, and we report increased resilience of large NNs to overtraining in case of multi-task training.

Sequence-to-sequence (seq2seq) approach for low-resource ASR is a relatively new direction in speech research. The approach benefits by performing model training without using lexicon and alignments. However, this poses a new problem of requiring more data compared to conventional DNN-HMM systems. In this work, we attempt to use data from 10 BABEL languages to build a multilingual seq2seq model as a prior model, and then port them towards 4 other BABEL languages using transfer learning approach. We also explore different architectures for improving the prior multilingual seq2seq model. The paper also discusses the effect of integrating a recurrent neural network language model (RNNLM) with a seq2seq model during decoding. Experimental results show that the transfer learning approach from the multilingual model shows substantial gains over monolingual models across all 4 BABEL languages. Incorporating an RNNLM also brings significant improvements in terms of %WER, and achieves recognition performance comparable to the models trained with twice more training data.

This paper presents a series of experiments we performed during our work on the MGB-3 evaluations. We both describe the submitted system, as well as the post-evaluation analysis. Our initial BLSTM-HMM system was trained on 250 hours of MGB-2 data (Al-Jazeera), it was adapted with 5 hours of Egyptian data (YouTube). We included such techniques as diarization, n-gram language model adaptation, speed perturbation of the adaptation data, and the use of all 4 ‘correct’ references. The 4 references were either used for supervision with a ‘confusion network’, or we included each sentence 4x with the transcripts from all the annotators. Then, it was also helpful to blend the augmented MGB-3 adaptation data with 15 hours of MGB-2 data. Although we did not rank with our single system among the best teams in the evaluations, we believe that our analysis will be highly interesting not only for the other MGB-3 challenge participants.

Training deep recurrent neural network (RNN) architectures is complicated due to the increased network complexity. This disrupts the learning of higher order abstracts using deep RNN. In case of feed-forward networks training deep structures is simple and faster while learning long-term temporal information is not possible. In this paper we propose a residual memory neural network (RMN) architecture to model short-time dependencies using deep feed-forward layers having residual and time delayed connections. The residual connection paves way to construct deeper networks by enabling unhindered flow of gradients and the time delay units capture temporal information with shared weights. The number of layers in RMN signifies both the hierarchical processing depth and temporal depth. The computational complexity in training RMN is significantly less when compared to deep recurrent networks. RMN is further extended as bi-directional RMN (BRMN) to capture both past and future information. Experimental analysis is done on AMI corpus to substantiate the capability of RMN in learning long-term information and hierarchical information. Recognition performance of RMN trained with 300 hours of Switchboard corpus is compared with various state-of-the-art LVCSR systems. The results indicate that RMN and BRMN gains 6 % and 3.8 % relative improvement over LSTM and BLSTM networks.

This paper provides an extensive summary of BUT 2016 system for the last IARPA Babel evaluations. It concentrates on multi-lingual training of both deep neural network (DNN)-based feature extraction and acoustic models including multilingual training of bidirectional Long Short Term memory networks. Next, two low-dimensional vector approaches to speaker adaptation are investigated: i-vectors and sequence-summarizing neural networks (SSNN). The results provided on three Babel Year 4 languages show clear advantage of both approaches in case limited amount of training data is available. The time necessary for the development of a new system is addressed too, as some of the investigated techniques do not require extensive re-training of the whole system.

Recent advancement in deep neural network (DNN) has surpassed the conventional hidden Markov model-Gaussian mixture model (HMM-GMM) framework due to its efficient training procedure. Providing better phonetic context information in the input gives improved performance for DNN. The state projection vectors (state specific vectors) in subspace Gaussian mixture model (SGMM) captures the phonetic information in low dimensional vector space. In this paper, we propose to use state specific vectors of SGMM as features thereby providing additional phonetic information for the DNN framework. To each observation vector in the train data, the corresponding state specific vectors of SGMM are aligned to form the state specific vector feature set. Linear discriminant analysis (LDA) feature set are formed by applying LDA to the training data. Since bottleneck features are efficient in extracting useful discriminative information for the phonemes, LDA feature set and state specific vector feature set are converted to bottleneck features. These bottleneck features of both feature sets act as input features to train a single DNN framework. Relative improvement of 8.8% for TIMIT database (core test set) and 9.7% for WSJ corpus is obtained by using the state specific vector bottleneck feature set when compared to the DNN trained only with LDA bottleneck feature set. Also training Deep belief network - DNN (DBN-DNN) using the proposed feature set attains a WER of 20.46% on TIMIT core test set proving the effectiveness of our method. The state specific vectors while acting as features, provide additional useful information related to phoneme variation. Thus by combining it with LDA bottleneck features improved performance is obtained using the DNN framework.