Bandwidth Extension WaveNet for Bone-Conducted Speech Enhancement

摘要

Bone-conducted (BC) speech is immune to background noise, but sounds muffled due to the characteristic of human body channel. The existing enhancement methods about BC speech mainly focus on the enhancement of the magnitude spectrum and ignore processing the mismatched phase. Besides, most of the existing communication systems of BC speech are based on low sampling rate configuration and cannot meet the better speech quality demand of the wideband speech communication. In this paper, a novel waveform generation method based on Bandwidth Extension WaveNet (BE-WaveNet) model is proposed, that builds the probability density between the conditional acoustic feature and the target waveform using the deep convolutional neural networks WaveNet and is able to generate the waveform from the enhanced magnitude spectrum directly. In order to further improve the speech quality, an up-sampling module of cross resampling rate is introduced, enabling the WaveNet to generate high-sampling rate speech with low-sampling rate acoustic feature, namely the BE-WaveNet is able to extend the bandwidth without increasing the communication cost. The experimental results show that compared with the waveform synthesis methods such as using the original phase and the phase estimation method of Griffm-Lim, the proposed method significantly improves the quality of BC speech. At the same time, high-sampling rate enhanced speech can be obtained without complex system configuration updates.