Study On Multi-Channel Dereverberation And Post-Processing Techniques

Distant speech signals captured in an enclosed space are inevitable affected by reverberation,which is caused by the reflections from walls and surrounding objects.The severe reverberation will deteriorate speech quality and intelligibility,as well as degrade the performance of the automatic speech recognition system.Therefore,the study on effective dereverberation algorithms has become particularly attractive in recent years.Researchers have proposed a variety of speech dereverberation algorithms,such as speech enhancement approaches for dereverberation,beamforming techniques for dereverberation,and statistical model-based dereverberation algorithms,etc.One of the most commonly used methods is Multi-Channel Linear Prediction(MCLP)speech dereverberation algorithm.Hence,MCLP speech dereverberation and post-processing techniques are studied in this thesis.Organization of this thesis is summarized as follows.Firstly,it can be noticed that Recursive Least Squares(RLS)algorithm is theoretically numerically unstable.The prototype RLS based MCLP adaptive dereverberation algorithm is improved and a QR-decomposition Recursive Least Squares(QR-RLS)based MCLP adaptive dereverberation algorithm is proposed.It has the same dereverberation performance and better numerical stability as the prototype algorithm.In order to further suppress the residual reverberation in the improved speech signal after processing by the dereverberation algorithm,the Wiener filter is used as a post-processing module of the improved algorithm.Then,based on the QR-RLS algorithm,the p-TA-QR-LS based MCLP adaptive dereverberation algorithm is derived.This algorithm can choose different parameter p according to the designed switching mechanism and then switch between different algorithm,enabling the algorithm to achieve a good balance between convergence rate and computational complexity.The simulation results show that the two improved MCLP adaptive dereverberation algorithms have good dereverberation performance and stability under the different reverberation levels.Second,for the problem that the exponential decay model is too dependent on the reverberation time,a late reverberation estimation model that does not require reverberation time estimation is proposed.In order to further suppress residual reverberation signal,taking advantages of the beamforming techniques and the Modified Log-Spectral Amplitude(OM-LSA)algorithm,an improved new model based MCLP adaptive dereverberation and post-processing algorithm is proposed.The simulation results demonstrate that the improved dereverberation algorithm can eliminate the late reverberation of speech signals under the different reverberation levels,and it has better performance than the prototype MCLP adaptive dereverberation algorithm.Furthermore,the improved algorithm effectively avoids the overestimation of the late reverberation by the exponential decay model,which reduces the distortion of the speech signal.