Research On Distant Speech Enhancement Methods In Reverberant Environments

Speech is one of the primary means for human-human interactions and human-machine communication.In an enclosed room,apart from noise,reverberation due to reflections from the walls,ceiling and floor,is the main cause of speech quality and intelligibility degradation.Since the desired speech signal is convolved with the room impulse response,the signal received by micro-phones is a superposition of the direct signal and the reverberant components.Room reverberation leads to temporal and spectral smearing,which would distort both the envelope and fine structure of a speech signal.At the same time,the presence of noise and potential interference,as well as the non-stationary nature of the speech itself,make speech enhancement more challenging.This the-sis focuses on the distant speech enhancement methods for microphone acquisition in reverberant environments.Considering the limitations of fixed beamforming,especially the differential beamforming,in time-varying room acoustic environments,this thesis proposes a distant speech enhancement frame-work of transfer function-generalized sidelobe canceller(TF-GSC)and improved minima con-trolled recursive averaging(IMCRA)based on adaptive beamforming.The joint spatio-temporal framework dereverberates and denoises the reverberant speech,meanwhile,guarantees the robust-ness of the dereverberation method in the presence of noise.Without the spatial position of the sound source and the acoustical transfer function as a pri-ori,this thesis replaces the TF-GSC structure with the multichannel linear prediction(MCLP)blind dereverberation framework,and introduces the least absolute sum method,the least absolute shrink-age and selection operator(LASSO)method under the group sparsity assumption and the normal-ized delay linear prediction method based on log-likelihood maximization.This thesis utilizes IMCRA as the post filter of above blind dereverberation methods and analyzes the influence of in-coherent noise on both frameworks.Simulations and experiments show that the MCLP framework outpeforms the GSC structure in acoustic conditions with predominant reverberation.Post-filtering is a necessary complement to both speech enhancement frameworks.The thesis designs and builds a multi-channel distant speech acquisition and processing system to realize real-time speech acquisition,processing and transmission.Moreover,the system verifies the effectiveness of the proposed distant speech enhancement methods based on GSC and MCLP framework with real recorded data.