Research On Multichannel Speech Enhancement Algorithm Assisted By Bone Conduction Sensor

Speech enhancement plays an important role in speech communication and humancomputer interaction.Compared with traditional single-channel speech enhancement,multi-channel speech enhancement can perform spatial filtering and has a better noise suppression effect.Generalized sidelobe canceller is a flexible and efficient multi-channel speech enhancement algorithm.It contains three modules: fixed beamformer,blocking matrix and adaptive noise canceller.In the real environment,factors such as reverberation and low signal-to-noise ratio may cause the desired signal to leak into the output of the blocking matrix,which in turn causes the desired signal to be distorted in the subsequent adaptive noise cancellation stage.In addition,the generalized sidelobe canceller can not effectively eliminate incoherent noise and diffuse field noise,and post-filtering after it can enhance noise suppression.This thesis mainly studies the fusion algorithm of boneconduction speech and multi-channel speech enhancement and post-filtering technology.The main works are summarized as follows.Firstly,in order to alleviate the leakage of the desired signal in the blocking matrix and prevent the cancellation of the desired signal in the adaptive noise canceller,this thesis utilizes a bone-conduction sensor to control the coefficients update of the two adaptive filters.Bone-conduction sensor is relatively insensitive to the ambient noise compared to the conventional air-conduction microphone.Hence,bone-conduction speech can be analyzed to generate very accurate voice activity detection results even in a strong noise environment.Then this thesis uses the voice activity detection results in the adaptive blocking matrix and adaptive noise canceller.By using voice activity detection to control adaptive blocking matrix and combining voice activity detection with signal-to-interference ratio to control adaptive noise canceller,the noise reduction performance of the generalized sidelobe canceller has been significantly improved.It is verified by experiments that the proposed algorithm not only improves speech quality remarkably but also boosts speech intelligibility.Secondly,in order to further eliminate the residual noise and interference in the output signal of the generalized sidelobe canceller,this thesis improves a post-filtering algorithm based on the transient beam to reference ratio.The sub-band signal processing approach is extended to the post-filtering for reducing computational complexity and eliminating musical noise.The relation between the observed signal,generalized sidelobe canceller output,and the reference noise signal is exploited to differentiate non-stationary noise components from speech components.In order to further reduce the computational complexity,this thesis also simplified the method of obtaining the transient beam to reference ratio and the a priori speech absence probability.Experimental results verify the effectiveness and robustness of the proposed algorithm.