Research On Speech Enhancement Algorithm Based On Sparse Representation And Binary Mask Estimation

In the process of moden speech communication,due to the interference of various noises,the quality and intelligibility of speech are decreased.Therefore,it is very important to improve speech quality and intelligibility in speech communication system.The limitations of traditional speech enhancement methods and their abilities are limited to deal with non-stationary noise.In this paper,the single channel speech enhancement problem is studied and three speech enhancement algorithms are proposed,which are based on the rapid development of sparse representation and binary masking model in recent years.The main areas of work and innovations of this paper are as follows:1.Most speech enhancement algorithms only change the amplitude of noisy speech to achieve the purpose of speech enhancement,but they ignore the effect of phase spectrum on speech.The phase spectrum compensation algorithm achieves the purpose of speech enhancement only by changing the phase of noisy speech,but it neglects the effect of amplitude on speech.Therefore,a speech enhancement algorithm based on sparse low rank model and phase spectrum compensation is proposed.Firstly,the noisy speech signal is processed by sparse low rank model.Then,a normalized least mean square adaptive filtering algorithm is used to optimize the compensation factor in the phase spectrum compensation algorithm.Finally,the improved phase spectrum compensation algorithm is used for the second enhancement of sparse low-rank speech,and the enhanced speech is obtained.The proposed algorithm has obvious effect by comparing with the traditional method.2.In order to further improve the speech enhancement effect of sparse low-rank model and phase spectrum compensation,a speech enhancement algorithm based on Itakura-Saito non-negative robust principal component analysis and phase spectrum compensation is proposed.This algorithm is superior to sparse low rank model and phase spectrum compensation algorithm for speech enhancement.3.The supervised learning methods produce better masking estimation.However,the mismatch between the real sound and the trained sound leads to the deterioration of its performance.Therefore,a binary mask estimation model based on local signal-to-noise ratio constraints is proposed.First,the noisy speech signal is processed by binary masking model,and then enhanced speech is obtained under the constraint of local signal-to-noise ratio(SNR).Compared with the experimental results,the algorithm shows excellent performance in the case of low signal to noise ratio and high signal to noise ratio.