A Speech Enhancement Neural Network Architecture with SNR-Progressive Multi-Target Learning for Robust Speech Recognition
Nan Zhou, Jun Du, Yan-Hui Tu, Tian Gao, Chin‐Hui Lee · 2019
We present a pre-processing speech enhancement network architecture for noise-robust speech recognition by learning progressive multiple targets (PMTs). PMTs are represented by a series of progressive ratio masks (PRMs) and progressively enhanced log-power spectra (PELPS) targets at various layers based on different signal-to-noise-ratios (SNRs), attempting to make a tradeoff between reduced background noises and increased speech distortions. As a PMT implementation, long short-term memory (LSTM) is adopted at each network layer to progressively learn intermediate dual targets of both PRM and PELPS. Experiments on the CHiME-4 automatic speech recognition (ASR) task, when compared to unprocessed speech using multi-condition trained LSTM-based acoustic models without retraining, show that PRM-only as the learning target can achieve a relative word error rate (WER) reduction of 6.32% (from 27.68% to 25.93%) averaging over the RealData evaluation set, while conventional ideal ration masks severely degrade the ASR performance. Moreover, the proposed LSTM-based PMT network, with the best configuration, outperforms the PRM-only model, with a relative WER reduction of 13.31% (further down to 22.48 %) averaging over the same test set.