Study On Acoustic Echo Cancellation And Beamforming Speech Enhancement

Speech is the most natural interactive mode.It also makes smart speakers popular with people.It makes people's lives more convenient.The signal to noise ratio is relatively low when people are far away from the smart device.The speech recognition effect will be poor without speech enhancement.However it is difficult to design a narrow beam main-lobe due to the number of elements and the size of the array.Besides it is difficult to wake up the keyword when the device sounds without acoustic echo cancellation.The microphone array Beamforming technology can effectively enhance the speech signal and the speech distortion is small.In this thesis,the microphone array signal is used to estimate the acoustic transfer function ratio for speech separation,and then Winer filtering is used for noise reduction.In addition,acoustic echo cancellation is studied.The contribution of this thesis is summarized as follows:(1)This thesis first introduces related background and the domestic and foreign research status about speech enhancement.(2)This thesis studies an improved adaptive filtering based on the AEC module in WebRTC open source code.The PBFDAF algorithm eliminates the echo in real time.We research anti-gradient blasting,the NLP and the NLMS algorithm.The purpose is to remove the residual echo,so that the double-talk detection can be omitted in the algorithm.Experiments show that the acoustic echo cancellation algorithm can achieve good results.(3)We combine the LCMV beamforming and the multi-channel post-wiener filtering algorithm so as to speech separation and noise reduction with the minimum mean square error criterion.The LCMV algorithm is used to remove the interference signal to obtain the desired signal.The Wiener-post filter is used to remove the residual noise.The calculated post filter coefficient causes the distortion of the desired signal.A balance is achieved between small distortion and the less introduction of background noise.Experiments show that the method studied in this paper can achieve good results in far-field speech enhancement and improve speech recognition rate.