Research On Signal-channel Speech Enhancement Based On Transformer

The aim of speech enhancement is to remove noise form speech signals to improve the speech quality and/or intelligibility.Speech enhancement is widely used in mobile communication,medical,military monitoring and other fields.It is also the front end of the speech recognition system.This thesis mainly focuses on speech enhancement technology based on the Transformer network.The major works are as follows:Firstly,a novel speech enhancement method,i.e.,ZTrans,which uses Transformer combined with a weighted residual mechanism was proposed.ZTrans can capture the longterm dependencies of the speech feature sequence and make full use of the relevant information between speech frames.In addition,a weighted residual mechanism was introduced into ZTrans to improve the model training speed.Experimental results show that the ZTrans method can obtain enhanced speech of high auditory speech quality.Secondly,a Complex-ZTrans model based on complex spectral mapping for speech enhancement task was proposed.It can conduct phase and amplitude enhancement simultaneously by predicting the real and imaginary components of the complex spectrum of speech.Experiments results show that the Complex-ZTrans model can obtain enhanced speech with higher speech quality than the ZTrans model.