Research On Frequency-Domain Blind Speech Dereverberation Algorithm

Voice signals generally exist in daily information interchange.During transmission,voice signals will mix with noise and reverberation generated by the outside world,resulting in waveform changes,reducing the intelligibility of received signals.Reverberation is caused by the multiple reflections of sound waves from hard objects when they travel in a relatively closed space.The reflected sound arrives at the receiver at different times.The reverberation effect is related to the attenuation rate of room impulse response.People pursue higher quality voice communication in complex environments,such as video conferencing,human-computer interaction and other scenes,so it is necessary to eliminate or suppress reverberation.Blind speech de reverberation consists of blind system identification and deconvolution.For the part of blind system identification,this article mainly discusses two methods of blind identification based on CR relation and eigenvalue decomposition and Kalman filter.After identifying the impulse response of the system,the inverse system can be obtained theoretically by using the MINT theorem and the observed signal can be filtered to achieve deconvolution.However,this method requires high identification accuracy,and small errors will lead to inaccurate deconvolution,so it is difficult to apply in practice.Kalman filtering deconvolution method has good noise stability and can overcome this problem,but its disadvantage is that it requires a lot of computation.First of all,in order to reduce the computation of deconvolution of Kalman filter,we propose a Kalman filter based blind SIMO system and its inverse system synchronization identification algorithm.This method can effectively reduce the computational complexity of the deconvolution part of the blind speech de reverberation algorithm on the premise of ensuring the noise stability of the algorithm.The computational complexity isO(L _h~2)reduced by one order of magnitude compared with the deconvolution algorithm based on Kalman filter.Secondly,according to the nature of cyclic convolution and the overlapping reservation method,the matrix operation with large amount of computation in the above time-domain algorithm can be realized in the frequency domain,so that the frequency domain algorithm for synchronous blind identification of SIMO system and its inverse system based on Kalman filtering can be obtained.The frequency domain algorithm includes two parts:(a)the inverse system identification part:the frequency domain calculation of the observation matrix obtained by the convolution of the identified system impulse response and the observed signal;(b)Inverse filtering part:use Fast Fourier Transform(FFT)to realize inverse filtering in frequency domain.In block operation,the average amount of computation per point is less than that of the corresponding time domain algorithm.Finally,a large number of simulation experiments are carried out for the above algorithms.The simulation experiment is carried out for the Kalman filtering de reverberation algorithm based on eigenvalue decomposition,and the effect of the algorithm is verified by comparing the impulse response waveform before and after identification and the NPM curve;Under different noises,the blind identification and deconvolution integration method of SIMO inverse system based on Kalman filter are simulated.Experiments show that the computational complexity of this algorithm is significantly reduced compared with the blind identification and deconvolution integration method of SIMO inverse system based on Kalman filter;The simulation experiment of the frequency domain implementation algorithm of SIMO inverse system blind identification and deconvolution integration based on Kalman filtering is carried out,and the calculation amount of the complex matrix calculation part of the two algorithms is counted.The calculation amount of the average point in each block of the frequency domain algorithm is lower than that in the time domain,and the calculation duration of the two algorithms is compared.Especially in the case of large room impulse response,the time consumption of the frequency domain algorithm is significantly less than that of the time domain algorithm,The frequency domain algorithm further reduces the computational complexity.