The Study On The Dual-Channel Speech Enhancement Technology In The Car Environment

With the rapid development of China's economy and the people's increasing consumption ability,automobile has become a major transportation vehicle.Compared with the traditional vehicle human-computer interaction mode based on touch-screen interaction,vehicle speech interaction is a safer and more intelligent way.In an ideal situation,the driver utters a speech command in order to control the in-vehicle electronic device,and the in-vehicle speech recognition system invokes the corresponding device based on the recognition results.However,there are various vehicle background noises and interferences from co-pilot position in the real vehicle environment,which seriously deteriorates the vehicle speech interaction experience.Therefore,the speech enhancement algorithm as a pre-processing scheme becomes an effective and necessary method to suppress noise and interference to facilitate subsequent speech interactions.Compared with the traditional single-channel speech enhancement algorithm,the microphone array algorithm can utilize the spatial information of the signal and has better noise suppression effect.This thesis mainly studies the dual-channel speech enhancement technique in the vehicle environment.The main work and contributions can be summarized as follows:Firstly,aiming at the problem that traditional dual-channel speech enhancement algorithms are not suitable for the car acoustic environment,this thesis firstly estimates the posterior probability that the signal time-frequency point is dominated by the target sound source or the interference sound source based on the way of observing the crosscorrelation power spectrum phase of the signal to suppress the influence of interference from co-pilot position.Then,the speech presence probability based on the deep neural network estimation is further combined to improve the robustness to the vehicle background noises.Finally,the mixed voice activity detection information is applied to the design of modules such as beamforming,blocking matrix,and adaptive noise canceller.Secondly,traditional post-filter algorithms can not effectively eliminate the residual co-pilot interference and vehicle background noises in the beamforming output of the vehicle environment.The performance of the noise power spectrum estimation algorithms in the classic single-channel and multi-channel post-filter algorithms are analyzed in depth and then an improved algorithm based on the variable leakage factor is proposed to improve the performance of the post-filter in the vehicle environment for the particularity of vehicle environment.Finally,computer simulation experiments based on the real vehicle environment data have been conducted and the simulation results show that the adaptive beamforming algorithm based on phase and the depth neural network hybrid voice activity detector and the post-filter algorithm based on the variable leakage factor noise power spectrum estimation can effectively reduce the distortion of the desired signal and suppress vehicle background noises and co-pilot interference,and the better noise reduction performance can be achieved.