Research On Single-channel Speech Separation Technology Based On Deep Learning

Single channel speech separation(SCSS)aims to recover several speech signals from 1-dimensional mixed speech.It is an underdetermined problem,which is extremely difficult to solve.SCSS is widely used in various fields due to its simple deployment and competitive prices.Therefore,the study of SCSS has a theoretical and practical significance.Recently,the deep neural network(DNN)has been exploited for the single channel speech separation problem.And DNN achieved outstanding performance by learning the nonlinear relationship between the mixed signal and the target sources.In this paper,we use deep neural network to solve the problem of single channel speech separation with or without background noise,and the proposed method optimizes the state-of-the-art approaches based on DNN.For single channel speech separation under noiseless and noisy conditions,a joint constraint algorithm and the amplitude-phase joint estimation algorithm are proposed accordingly.The results demonstrate that the proposed algorithms have great performance in speech separation.The contributions of this paper are summarized as follows:(1)The significance and development of speech separation are introduced.The principle and training process of deep neural network are explained in detail,and then the acoustic features and training targets are described.Besides,the system framework and the procedure of realization in terms of single channel speech separation based on deep neural network are analyzed.(2)To solve the problem of single channel speech separation in noiseless environment,a joint constraint algorithm of single channel speech separation based on dual output deep neural network is proposed.Based on the traditional method,the deep neural network is trained under the guidance of the new loss function with joint constraints.The proposed loss function fully exploits the joint relationship between the dual outputs of DNN,and further constrains the parameters of the network to make the estimated speech amplitude spectrum approach the pure speech amplitude spectrum as much as possible.The simulation results show that the proposed algorithm can improve speech intelligibility and reduce the distortion of speech signal effectively.(3)To solve the problem of single channel speech separation in noisy environment,a joint estimation algorithm of amplitude and phase based on DNN is proposed.Traditional methods train a DNN to separate the noisy mixed signal directly,but ignore the influence of noise and phase on the separation performance.Therefore,an algorithm to estimate the amplitude and phase of speech signal simultaneously is proposed.In terms of amplitude estimation,the mixed signal with noise is denoised by enhanced DNN,and then the amplitude spectrum after denoising is separated by amplitude separation DNN.In terms of phase estimation,DNN is used to estimate the phase of the target speaker.Finally,the speech signal can be reconstructed by inverse short-time Fourier transform.The experimental results verify the effectiveness of the proposed algorithm.