面向舱室声学环境的深度时域语音增强网络

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.