No abstract

No abstract

Data augmentation is one of the most effective ways to make end-to-end automatic speech recognition (ASR) perform close to the conventional hybrid approach, especially when dealing with low-resource tasks. Using recent advances in speech synthesis (text-to-speech, or TTS), we build our TTS system on an ASR training database and then extend the data with synthesized speech to train a recognition model. We argue that, when the training data amount is low, this approach can allow an end-to-end model to reach hybrid systems' quality. For an artificial low-resource setup, we compare the proposed augmentation with the semi-supervised learning technique. We also investigate the influence of vocoder usage on final ASR performance by comparing Griffin-Lim algorithm with our modified LPCNet. An external language model allows our approach to reach the quality of a comparable supervised setup and outperform a semi-supervised setup (both on test-clean). We establish a state-of-the-art result for end-to-end ASR trained on Lib-riSpeech train-clean-100 set with WER 4.3% on test-clean and 13.5% on test-other.

Mixup is a recently proposed technique that creates virtual training examples by combining existing ones. It has been successfully used in various machine learning tasks. This paper focuses on applying mixup to automatic speech recognition (ASR). More specifically, several strategies for acoustic model training are investigated, including both conventional cross-entropy and novel lattice-free MMI models. Considering mixup as a method of data augmentation as well as regularization, we compare it with widely used speed perturbation and dropout techniques. Experiments on Switchboard-1, AMI and TED-LIUM datasets shows consistent improvement of word error rate up to 13% relative. Moreover, mixup is found to be particularly effective on test data mismatched to the training data.

The goal of this work is to build a state-of-the-art English conversational telephone speech recognition system. We investigated several techniques to improve acoustic modeling, namely speaker-dependent bottleneck features, deep Bidirectional Long Short-Term Memory (BLSTM) recurrent neural networks, data augmentation and score fusion of DNN and BLSTM models. Training set consisted of the 300 hour Switchboard English speech corpus. We also examined the hypothesis rescoring using language models based on recurrent neural networks. The resulting system achieves a word error rate of 7.8% on the Switchboard part of the HUB5 2000 evaluation set which is the competitive result.

No abstract

No abstract

This paper describes the Speech Technology Center (STC) system for the 5th CHiME challenge. This challenge considers the problem of distant multi-microphone conversational speech recognition in everyday home environments. Our efforts were focused on the single-array track, however, we participated in the multiple-array track as well. The system is in the ranking A of the challenge: acoustic models remain frame-level tied phonetic targets, lexicon and language model are not changed compared to the conventional ASR baseline. Our system employs a combination of 4 acoustic models based on convolutional and recurrent neural networks. Speaker adaptation with target speaker masks and multi-channel speaker-aware acoustic model with neural network beamforming are two major features of the system. Moreover, various techniques for improving acoustic models are applied, including array synchronization, data cleanup, alignment transfer, mixup, speed perturbation data augmentation, room simulation, and backstitch training. Our system scored 3rd in the single-array track with Word Error Rate (WER) of 55.5% and 4th in the multiple-array track with WER of 55.6% on the evaluation data, achieving a substantial improvement over the baseline system.