Improving Deep Neural Network Based Speech Enhancement in Low SNR Environments

摘要

We propose a joint framework combining speech enhancement (SE) and voice activity detection (VAD) to increase the speech intelligibility in low signal-noise-ratio (SNR) environments. Deep Neural Networks (DNN) have recently been successfully adopted as a regression model in SE. Nonetheless, the performance in harsh environments is not always satisfactory because the noise energy is often dominating in certain speech segments causing speech distortion. Based on the analysis of SNR information at the frame level in the training set, our approach consists of two steps, namely: (1) a DNN-based VAD model is trained to generate frame-level speechon-speech probabilities; and (2) the final enhanced speech features are obtained by a weighted sum of the estimated clean speech features processed by incorporating VAD information. Experimental results demonstrate that the proposed SE approach effectively improves short-time objective intelligibility (STOI) by 0.161 and perceptual evaluation of speech quality (PESQ) by 0.333 over the already-good SE baseline systems at-5dB SNR of babble noise.