Dereverberation Based On Subband Spectrum Substraction And Microphone Array

In enclosed space, direct arrival acoustic signal with reflect acoustic signal which reflected by wall?ceiling and ground propagate to ear or microphone, but reflect acoustic signal later than direct arrival acoustic signal and form a long tail which dim clean acoustic signal. So the phenomenon makes human communication and voice system hard. With the development of science and technology, speech enhancement technology is more closely linked with many fields, in the paper, the following aspects are studied:(1) The effect of convention Spectral Subtraction dereverberation is studied.Aiming the characteristic that different bands of reverberation signal have different reverberation time. So we use Spectral Subtraction on each bands to improve dereverberation precision and assess the impression dereverberation signals respectively come from Spectral Subtraction and Subband Spectral Subtraction with Spectrogram and Perceptual Evaluation of Speech Quality. The results show that Subband Spectral Subtraction is better.(2) The technique of Delay-and-Sum dereverberation is studied, and applied to Microphone Array with four sensors. Paper estimates the time delay between channels with different method: PHAT-DS?PBF-DS?PBF-PHAT-DS,then align and sum different channels signal to make the main flap of Microphone Array aims at the angle of direct arrival signal. Different scenes experiments of audio acquisition have carried on. Paper also introduces the experimental process and specification reduce the error during the experiment. Then the effects of three algorithm are compared, the results show that PBF-PHAT-DS is better.(3) The Linear Prediction Technique is applied to dereverberation, considering that change rate of impulse response in enclosed space is far slower than vocal tract,Paper designs a kind of equalizer based on the minimum mean square error criterion which aim at enclosed space reverberation system.considering the characteristics that speech recognition system is sensitive to the middle and late reverberation,Paper designs weighted least mean square error equalizer. The simulation results show that the weighted least mean square error equalizer can significantly suppress the middle and late reverberation energy.