The STC System for the CHiME 2018 Challenge

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.

The STC ASR System for the VOiCES from a Distance Challenge 2019

“…i-vectors, is a widely used technique [35]. Our CHiME5 experience [36] revealed that speaker adaptation is extremely useful not only for telephone and microphone speech recognition tasks but for distant ASR as well. Based on this experience, we decided to apply modern x-vectors speaker embeddings [16,17] instead of i-vectors for speaker adaptation.…”

Improved Speaker-Dependent Separation for CHiME-5 Challenge

“…Although it's inferior to the USTC-iFlytek's, our system perform separation only once and has low computational complexity and model size apparently. The details on each ses- [32] 62.09 Toshiba [33] 63.30 STC [23] 63.30 RWTH-Paderborn [34] 68.40 Official [13] 80.28 sion and location over official AM and ours are given in Table 6. Even based on the official backend, our SD separation frontend contributes a 10% WER reduction, which is a significant improvement on this challenging task.…”

“…In [14], suffering from low-quality training targets, the system just achieved 2% absolute reduction on WER. Second, inspired by [21,22,23], we incorporate i-vectors as auxiliary features, which aims to extract the target speaker. With the speaker-aware training technique, we achieve much better results using only one mask estimation model.…”

CNN-based Multichannel End-to-End Speech Recognition for Everyday Home Environments

“…overlapped sentences or unfinished utterances, noises from home appliances at a signal-to-noise ratio (SNR) between 5 and 20 dB, distant microphone speech, and a small training dataset of 40 h (i.e., low resources). Most competitive systems, except for [12], in the fifth CHiME challenge employ conventional ASR methods with multichannel speech enhancement techniques [15], [16], [17], [18].…”