Research And Implementation Of Speech Dereverberation Algorithms

Reverberation is a common acoustic phenomenon,which can be seen when recording indoors such as live video broadcasts,auditorium speeches,and round table conferences.For many years,reverberation has been a double-edged sword.In the field of music processing,reverberation is like magic,making the sound ethereal,thick and melodious.But in the field of speech signal processing,reverberation is destruction.The culprit of signal intelligibility and clarity seriously reduces the quality of speech communication and recognition accuracy,and affects the performance of speech equipment.The speech dereverberation algorithm,as a speech front-end preprocessing technology,plays a key role in reducing reverberation and improving the quality of the speech spectrum,therefore it has important value in term of research.Based on the existing various speech dereverberation technologies,the research of the thesis focuses on the multi-channel linear prediction(MCLP)dereverberation algorithm because of its better performance and strong universality currently.It is also called the weighted prediction error algorithm.The MCLP doesn't rely on the prior knowledge of reverberation environment and has no requirement for microphone array type,but the existing MCLP algorithm also has some shortcomings such as a rough way of estimating the expected signal and inefficient in covering the reverberation time by the given the order of prediction coefficients.In response to the existing problems,this thesis has made the following improvements:(1)An improved scheme based on the estimation of desired signal power spectral density(PSD).The existing MCLP algorithm directly uses the observation signal when estimating the PSD of the desired signal,this leads to inaccurate prediction of the reverberation component,and reduces the stability speed of the algorithm in the early stage,and affects the auditory feelings.This thesis combines a reverberation component estimation method with a geometric spectral subtraction technique,and designs an improved scheme based on the desired signal PSD estimation.First,the exponential decay model and the calculated reverberation time are used to estimate the reverberation PSD of the current frame,and then the desired PSD through geometric spectral subtraction is obtained to improve the accuracy of the estimation and avoid the problem of over-subtraction caused by conventional spectral subtraction.The test results show that the improved MCLP algorithm based on the desired signal PSD estimation can stabilize the dereverberation effect more quickly,and is better than the original algorithm in all speech quality evaluation indicators.(2)An improved scheme based on prediction error compensation.This thesis studies the effects of different speech channel numbers and linear prediction coefficient orders on the performance and computational cost of the MCLP algorithm.It is found that compared to increasing the number of channels,increasing the prediction order can be more effective while increasing the same amount of computational complexity.On the basis of the improved scheme for PSD estimation,a prediction error compensation method is proposed to solve the problem that the prediction order is difficult to cover all the signals in the reverberation time within the allowable complexity range.The predicted value within the range is regarded as the linear prediction error and the current output is used to characterize the error amount.The error is minimized to get the error gain and then get the prediction coefficient and the expected signal after gain compensation.The test results show that the prediction error compensation scheme can further remove the reverberation residue.In terms of algorithm engineering implementation,this thesis uses C language and scientific computing library to compile an efficient engineering version of the improved scheme,designs a speech streaming scheme,and implements real-time dereverberation processing on embedded devices.