Study On Speech Enhancement With Reverberation

With the rapid development and application of speech communication and speechbased human-machine interaction system,the technology of speech enhancement has to deal with increasingly more complex environment.Reverberation,existing in any enclosure space,always degrades the quality and intelligibility of speech,and is also a big challenge to the blind source separation system and the automatic speech recognition system.It is of great importance to improve the performance of speech enhancement system in the reverberant environment.In this thesis,speech enhancement with reverberation is investigated with focuses on online adaptive dereverberation,online speech separation with reverberation,and speech enhancement based on deep neural network.The adaptive algorithm based on multi-channel linear prediction(MCLP)is an effective dereverberation method.However,the abrupt change of target speaker position always forms an obstacle to the adaptive algorithm.In this thesis,the recursive least squares(RLS)-based MCLP algorithm and the Kalman filter-based MCLP algorithm are investigated.Based on the relative change of adaptive filter coefficients,a time-varying forgetting factor is designed for RLS algorithm and a re-initialization mechanism is proposed for Kalman filter to improve the robustness of the algorithms to the abrupt change of target speaker position.The advantages of the proposed scheme are demonstrated in the experiments.Multi-microphone blind source separation(BSS)is an effective method to extract the target speech signal when several speech sources coexist and can be utilized as an important part of a speech enhancement system.However,the performance of BSS algorithm always degrades in reverberant environments.This paper focuses on implementing BSS with reverberation on a miniature dual-microphone system.The beamforming based on the first-order differential microphone array is utilized to acquire the speech signal.To further improve the performance of speech separation in highly reverberant environments,the dereverberation based on multi-channel linear prediction is also combined into the system and a joint optimization algorithm of speech separation and dereverberation is proposed.The benefits of the proposed system are demonstrated by experiments.Deep neural network(DNN)-based speech enhancement method has shown its appealing performance in certain scenarios.Time-domain audio separation network(TasNet)is a time-domain end to end network that can effectively separate speech and noise.In this thesis,the performance of the fully convolutional TasNet(Conv-TasNet)is utilized in low signal-to-noise(SNR)reverberant situation.A network combining speech separation and dereverberation is proposed,and the strategies to improve online processing are also discussed.The experiments show the benefits of the proposed methods.