Improvement Of Multi-channel Dereverberation And Noise Reduction Algorithm Based On EM

Speech noise reduction and dereverberation are important parts of the front-end acoustics processing technology in speech recognition systems.For application scenarios such as intelligent conference transfer system,voice-activated home operating system,and robot assistant,in order to improve the quality of speech obtained and achieve acoustic signals that meet the requirements of speech recognition in the case of high-noise,reverberation,and far distance sound sources,the microphone array is usually used for voice processing.Therefore,the noise reduction and dereverberation based on microphone arrays in far-field conditions are research hotspots in speech processing technology.Multi-channel speech noise reduction and dereverberation utilize spatial signals in different spatial directions collected by microphone groups placed in a certain geometric structure(usually linear and annular)to perform space-time processing to achieve noise suppression and dereverberation,thereby improving speech signal processing quality and speech recognition rate in the real environment.The common techniques for implementing combined multi-channel speech noise reduction and dereverberation include spectrum enhancement techniques,probability-based model techniques and acoustic multi-channel equalization techniques.The expectation maximizaition(EM)algorithm has been applied by many people to speech dereverberation technology.In the investigation and analysis of the multi-channel dereverberation and noise reduction algorithm based on EM,it was found that previous scholars either replaced the late reverberation directly with an ideal diffuse sound field or directly estimated the noise as a known variable by estimating received speech.In order to improve the ability of the algorithm to reduce noise and dereverberation,we difine the noise difference variable.The noise difference variable is also set as a hidden variable,and the estimated noise is used as the initial value of the algorithm iteration,which is solved by EM iteration.In order to further improve the applicability of the algorithm and solve the problem of dereverberation and noise reduction of speech in high noise conditions,utilizing the advantage that the original algorithm performs well under the high signal-to-noise condition,we combine the original EM algorithm with spectrum enhancement technology.First using a minimum variance distortionless response(MVDR)beamformer(BF)to complete noise reduction,and then performing dereverberation with EM algorithm.Under almost noiseless conditions,in order to prevent the phenomenon that the EM algorithm does not converge or converge too slowly,only the silenced speech is used as a hidden variable.The above two ideas were tested and verified by simulation experiments.The evaluation indicators were Perceptual Evaluation of Speech Quality(PESQ),Log-Spectral Distance(LSD)and segmented signal-to-noise ratio(SNR).Experiments show that under high SNR conditions,setting the noise difference variable as a hidden variable can obtain a higher PESQ score and a smaller LSD value;under low SNR conditions,the EM algorithm combined with spectral enhancement techniques can obtain better PESQ score and higher segmented signal to noise ratio.