The Research Of Speech Enhancement Methods Based On Soft Masking Model And Phase Spectrum Compensation

Speech enhancement is a digital signal processing technology,which tries to obtain the clean speech from noisy speech as much as possible.Although many speech-related applications can work well in a quiet environment,their performance would suffer serious decline when a lot of noise exist.Most of existing speech enhancement methods can improve speech quality effectively,as for the speech intelligibility,the improvement is very limited.In addition,many enhancement methods would increase complexity of the process to pursue a better effect which leads to a poor real-time performance,thus makes those methods difficult to apply.This paper gives a detailed analysis of the soft masking estimator based on a posteriori signal to noise ratio(SNR)uncertainty(SMPO).The SMPO method combines the ideal binary mask(IdBM),thus it can effectively improve the speech quality and intelligibility.Meanwhile,the paper points out that the SMPO has a shortcoming in noise removal.On this basis,the paper proposed two improved methods.The first method selects two common ways and applies the noise removal on speech segments whose SNR value is higher than a threshold.The second method takes the auditory masking effects into consideration,which uses a accurate way to processes the speech segments whose SNR value is between a small threshold and a high threshold.The simulation results show that both improved methods achieve good enhancement.Between them,the second improved method performs better.Phase spectrum is important for speech enhancement.The phase spectrum compensation(PSC)method has a good enhancement effect and its enhancement process is very simple.But this method has a problem in parameter setting.In the improved method,the noisy signal is divided into speech segments and noise segments using existing parameters and then processed separately.In this way,the constant denoising factor is modified to adaptive parameter.As a result,the improved method outperforms the original method and maintains good real-time performance.Apart from that,this paper combines the improved method with the methods based on magnitude spectrum modification.The combined methods process phase spectrum and magnitude specturm separately.It turns out the combined methods have good performance and have an obvious superiority compared with the methods based on magnitude spectrum modification.In short,the simulation results show that all the improved methods outperform the original methods.