Convolutional Neaural Network For Speech Separation

With the gradual popularization of smart phones and smart devices,the technology of voice interaction between human and machine has received widespread attentions.To make the interaction convenient and efficient just like interaction among people is one of the hottest topics of research in last few years.However,one important problem is speech separation,which has a great effect on the accuracy of automatic speech recognition,voice automatic translation and speaker recognition.In recent years,with the development of the deep learning theory,the speech separation based on deep model has gradually become a new research trend,and gets a good performance.However,the research of speech separation based on Convolutional Neural Network(CNN)is few.Compared to Deep Neural Network(DNN),CNN performs well on depicting temporal and spatial characteristics of input speech signals,and characterizes speech features better.So,in this paper,we propose to use convolution neural networks to deal with speech separation.This paper has two experiments of single channel speech separation and multi-channel speech separation.Each experiment is implemented on the same data set using CNN and DNN.The experiment of single channel speech separation is conducted on the standard speech database TIMIT.While,Multi-channel speech separation is evaluated on the official data of CHIME3.Experimental results show that compared with deep neural network convolution neural network,CNN can not only attain higher Perceptual Evaluation of Speech Quality(PESQ)and Short Time Objective Intelligibility(STOI),but also reduce the model complexity and the training weight parameters.