Research On Single-microphone Speech Dereverberation Algorithm

In real life,speech is always disturbed by various reverberations.Espectially in confined indoor spaces,when we use hands-free telephones,teleconferences and other occasions,reverberation occurs when the sound source is located far from the sound receiver.Reverberation can reduce the clarity and intelligibility of speech and seriously affect people's hearing experience.Speech dereverberation is an important part of speech enhancement.It provides preprocessing for speech signal processing,which can be classified as speech synthesis,sound source localization and speech recognition.Its performance directly determines the effect of speech enhancement.Therefore,speech dereverberation technology has a wide range of research foundations and very important application values.In this dissertation,the commonly used single-microphone speech dereverberation algorithm is systematically studied.And the two-stage dereverberation algorithm and the spectral enhancement algorithm based on Hidden Markov Model are studied respectively.In the process of studying the two-stage dereverberation algorithm,it is found that the two-stage dereverberation algorithm has a better dereverberation effect when the reverberation time is less than 0.4s.However,when the reverberation time increases,its dereverberation performance is seriously degraded.At the same time,the background noise cannot be well suppressed.To solve this problem,this paper deeply studied the linear prediction technology and Gammatone filter and designed an improved two-stage dereverberation algorithm.In the process of studying the spectral enhancement algorithm based on Hidden Markov Model,it is found that the existing spectral enhancement algorithm based on Hidden Markov Model can effectively remove late reverberation and background noise,while it has limited effect on early reverberation.In order to completely remove the influence of reverberation,this paper designs and implements a spectral enhancement algorithm based on linear prediction and Hidden Markov Model.(1)Improved two-stage algorithm.By studying and analyzing reverberation speech's characteristics,it's find that the reverberation speech has different reverberation times in different frequency bands.And the length of the inverse filter is related to the reverberation time.Accurate estimation of the inverse filter can achieve good results in a long-term reverberation environment.In this paper,a Gammatone filter is used to divide the reverberation speech subbands according to the human auditory model.Then the Schroeder reverberation time estimation method is used in each subband to obtain the reverberation time under different frequency bands.After that,adaptive selection of inverse filter length is choosed according to the reverberation time.Experiments show that the maximum residual skewness based on linear prediction is more robust to noise than the inverse filter based on maximum kurtosis estimation.Finally,in each subband,an inverse filter of different lengths is estimated based on the residual skewness,so as to improve the performance of the dereverberation algorithm.(2)Spectral enhancement algorithm based on linear prediction and hidden markov model.Based on the research of linear predictive inverse filter technology and the existing Hidden Markov Model based spectral enhancement method,the early reverberation is removed by estimating the inverse filter to preprocess the reverberation speech.Then the Taylor vector series is used to update the reverberation parameters adaptively to improve the accuracy of the spectrum enhancement function.In this paper,the two-stage dereverberation algorithm,the spectral enhancement method based on Hidden Markov Model and the two improved algorithms are simulated on MATLAB platform and their performances are compared.Experimental results show that the improved two-stage dereverberation algorithm outperforms the existing two-stage algorithm both in the dereverberation effect and the suppression of the environmental noise.The spectral enhancement algorithm based on linear prediction and Hidden Markov Model is better at the same reverberation time than the existing spectral enhancement algorithm based on Hidden Markov Model.The shorter the reverberation time is,the better the performance improvement is.