Single Channel Speech Separation Based On Deep Learning

Speech separation includes both the speech separation of multiple speakers,and the separation of speech and noise.The main work of this paper focuses on the separation of speech and noise which is generally called speech enhancement.With the development of artificial intelligence,the voice interaction is becoming more and more widely in reality.The interference of noise seriously restricts the performance of voice interaction.So,the separation of speech and noise is of great importance.Since many of the scene of voice interaction is based on single microphone,the speech separation technology based on single microphone is paid more and more researchers' attention.The traditional single-channel speech separation algorithm can be divided into two categories:unsupervised and supervised single-channel speech separation.Unsupervised single channel speech separation is mostly based on digital signal processing technology,such as spectral subtraction,Wiener filtering,etc.Commonly used supervised speech separation algorithms are based on shallow artificial neural network,Non-negative Matrix Factorization(NMF)and Hidden Markov Model(HMM).With the development of Deep Neural Network(DNN)technology in recent years,the single-channel speech separation based on DNN has made great progress.The strong nonlinear learning ability of DNN makes the speech separation perform well,and it has gradually become a new trend in speech separation task.This paper first analyzes the advantages and disadvantages of the traditional speech separation algorithm and the DNN.Then,two algorithms are proposed:(1)A joint optimization model based on DNN and Non-negative Matrix Factorization(NMF).(2)A joint optimization model based on DNN and Convolutive Non-negative Matrix Factorization(CNMF).Finally,a series of experiments were conducted to test the efficiency of the proposed methods.