面向舱室声学环境的深度时域语音增强网络

Single-channel deep time-domain speech enhancement networks for cabin environments

针对舱室环境单通道语音增强问题,设计了一种联合并行空洞卷积与分组卷积的深度时域语音增强网络。该网络以经典卷积时域音频分离网络为基础,在增强层设计中通过不同膨胀因子执行两路并行的空洞卷积操作,实现对长时信号的处理以更多地提取信号包络所包含的低频信息并抑制噪声混响所带来的时延问题,同时保留了局部的语音细节信息,提高对波形中所包含语音及背景噪声谐波信息的提取准确度;另外,利用分组卷积降低并行卷积操作所导致的网络规模扩大,使网络在具有良好增强效果的同时能够保持较小的网络规模及运算复杂度。以多类飞机舱室噪声为数据基础的实验表明,所设计的网络模块相较于基线网络提升了客观评价指标值,与现有其他常用网络的比较结果表明此方法在舱室环境的数据条件下可获得更好的主客观语音增强评价指标,且在高噪声级的线谱及窄带处具有更低的失真度。

A deep time-domain speech enhancement network with combined parallel dilated convolution and group convolution is designed for the single-channel speech enhancement problem in cabin environment.The network is proposed based on the classical convolutional time-domain audio separation network.In the enhancement layer,the parallel cavity convolution operations are performed with different expansion factors to realize the processing of long-time signals to extract more low-frequency information contained in the signal envelope and suppress the time delay problem caused by noise reverberation.Meanwhile,the speech detail information is preserved and the extraction accuracy of speech and background noise harmonic information contained in the waveform can be increased.In addition,group convolution is used to reduce the expansion of network size caused by parallel convolution operation,so that the network can maintain a small network size and operation complexity while having good enhancement effect.The experiments based on multiple types of aircraft cabin noise show that the designed network module improves the objective metrics compared with the baseline network,and the comparison results with other existing common networks show that the method can obtain better subjective and objective speech enhancement evaluation indexes under the data conditions of cabin environment,and has lower distortion in the line spectrum and narrow band of high noise level.