Adaptive Speech Dereverberation Based On Microphone Array

In a relatively closed space,there is reverberation effect,that is,when someone is speaking or the speaker is playing audio,there is reverberation — echo in the audio signal received by the listener or microphone.In the case of speech communication and human-computer interaction,reverberation effect will affect the clarity and intelligibility of speech signals.Speech dereverberation is a hot research field in speech signal processing.As far as the reverberation mechanism is concerned,speech dereverberation is essentially a blind deconvolution process,which includes blind system identification and deconvolution.Aiming at the problems existing in the blind deconvolution algorithm based on Kalman filtering,the optimization algorithm of kalman filtering deconvolution and synchronic blind multichannel identification and deconvolution with Kalman filter are proposed.The main work of this paper includes the following aspects:1.In view of the kalman filtering in the convolution algorithm,the problem of the large amount of calculation,the author puts forward a simplified algorithm,in the standard kalman filtering solution of convolution algorithm,kalman gain calculation is only related to the convolution of the system impulse response,and the convergence is fast,but the proportion is very high amount of calculation,and the state vector calculation is relatively simple,only related to the kalman gain and system input.Therefore,in the time period when the impulse response of the convolution system is constant,the kalman gain fast convergence can be used to calculate the convolution system separately,and then the state vector can be calculated after the convergence,which can effectively reduce the average calculation amount of the deconvolution process.2.This paper proposes a synchronous dereverberation algorithm based on Kalman filter.The state vector of Kalman filter is composed of multi-channel system parameters and source signal vectors,and the process equation and measurement equation are based on the input-output Relation of SIMO system and the Cross Relation between channels.In addition,the identification part and deconvolution part of the blind system can be decoupled to generate two seemingly independent Kalman filtering problems,and these two kalman filtering problems can be realized in parallel computation.3.A large number of simulation experiments have been done for the above two algorithms.As for the optimization of kalman filter deconvolution,the simulation experiment shows that the optimization algorithm sacrifices a small amount of convergence speed,but reduces a large amount of computation.The simulation results show that the parallel structure is more beneficial to algorithm optimization and real-time signal processing than the cascade structure.