Time-delay Neural Network For Speech Separation

With the rise of the age of artificial intelligence,human-machine voice interaction technology is infiltrating into all walks of life.Products with voice interaction technologies are also emerging.However,the sound signal in real life is always disturbed by noise,and the machine cannot actively track the target speech from the noisy sound like the human ear.So the speech separation technology is particularly important.Deep learning technology become more and more important in artificial intelligence.In recent years,researchers have begun to use deep learning to achieve the task of speech separation.This paper studies the speech separation based on Time Delay Neural Networks(TDNN).Compared with Deep Neural Networks(DNN),the TDNN model can make good use of the spatio-temporal correlation between frame-level features;and Convolutional Neural Networks(CNN).Two experiments of single channel speech separation are conducted in this article.For the TDNN model,different TDNN structures are generated under different offsets.Therefore,this thesis first analyzes the performance of TDNN with different structures in speech separation performance.Second,TDNN and DNN,CNN are evaluated and compared.The experimental results show that the TDNN structure with wider input layer context can achieve better speech separation effect;compared with DNN,the separation effect of TDNN is slightly improved;compared with CNN,the speed of TDNN model in training and testing is faster.