Research And Implementation Of Single-Channel Speech Enhancement Technology In Non-Stationary Noise Environment

In the past decades,speech enhancement technology has developed rapidly.From the application of traditional unsupervised methods to the rise of deep learning methods,various methods of speech enhancement have achieved a lot,but they still have unlimited potential.This research studied single-channel speech enhancement technology which is widely used nowadays.First of all,this paper gives the experimental verification of the traditional unsupervised single-channel speech enhancement method,including the spectrum subtraction,Wiener filtering method,and the Kalman filtering method,to evaluate the performance of speech enhancement under stationary noise and non-stationary noise environment.And analyzed them in a non-stationary noise environment which can't perform well.In addition,a single-channel speech enhancement method based on masking prediction network have been proposed.Two deep neural networks are constructed: the masking prediction network and the auxiliary network.To guide the training of the masking prediction network,the mixed noisy speech and the corresponding pure speech were used as the input of the masking prediction network and the auxiliary network,with the added attention function was used to generate the speech parameter information at every relevant moment.Finally,a single-channel speech enhancement method based on spectrum mapping network have been proposed.In this method,the frequency domain information of speech signal is used to realize the separation of speech spectrum and noise spectrum by cascading spectral mapping separation network and feature extraction network which add masking output layer.By embedding the attention mechanism into the deep neural network,the proposed approach has a stronger performance improvement effect in non-stationary noise environment.