Research On Time Domain Speech Enhancement Algorithms Based On Auditory Perceptual Weighting

A speech enhancement algorithm that utilizes human auditory masking characteristics can effectively improve noise suppression performance.Among them,the perceptual weighting filtering technique is applied in a variety of single-channel speech enhancement algorithms because it does not need to estimate the masking threshold and avoids the error caused by estimating the masking threshold.Compared with the transform domain,the time domain speech enhancement algorithm has no music noise and avoids the error caused by signal conversion.Therefore,this paper will discuss the application of perceptual weighting technology in time domain speech enhancement.However,in the time-domain single-channel speech enhancement,the influence of reverberation on perceptual weighting technology is not discussed.Therefore,the perceptually-based subspace approach is taken as an example to discuss the performance of perceptual weighting technique in reverberation environment.In addition,the perceptual weighting technique has not been extended to multi-channel speech enhancement algorithms in time domain.For this reason,this paper extends this technique to two time-domain multi-channel speech enhancement algorithms,space-time prediction and tradeoff filtering.The main work and contributions of this paper are as follows:1.The performance of time domain single channel speech enhancement algorithm based on auditory perception weighting in reverberation environment is studied.Firstly,the time domain wiener filtering is derived theoretically,and the perceptual weighting technique is introduced.It is concluded that the perceptual weighting technique is invalid under the condition of no residual noise.When there is residual noise limitation,it is concluded that the perceptual weighting filter can improve the performance of noise reduction.Then the subspace approach based on time domain constraint estimator and the perceptually-based subspace approach are derived respectively.Finally,taking the perceptually-based subspace approach as an example,in the room reverberation environment,the enhancement effect of the algorithm under different SNR and different noise types is explored.When the speech distortion is roughly equal,the output SNR and PESQ scores of the perceptually-based subspace approach are improved compared with the original method,which shows that the perception weighting technology enhances under reverberation conditions.Performance improvement is effective.2.The perceptual weighting technique is extended to the time domain multi-channel speech enhancement algorithm.The perceptually-based space-time prediction approach and the perceptually-based tradeoff approach are proposed.The space-time prediction approach and tradeoff approach are deduced theoretically,and it is found that under the condition that the estimated signal is completely undistorted,the perceptual weighting technique can not improve the noise reduction performance of the two algorithms.When a certain distortion is allowed,the perceptual weighting technique can improve the enhancement performance of the two algorithms.The simulation experiments under different conditions verifies the effectiveness of perceptual weighting technique in time domain multi-channel speech enhancement algorithm.In addition,as the speech distortion increases,the enhancement effect is more significant.