Single-channel Speech Enhancement Method Based On Binuaral Cues

Speech enhancement has been widely investigated with the improvement of the communication technology.Although some traditional method,such as spectral-subtraction method,Wiener filtering method,and statistical-model-based methods and so on,have obtained a good performance for stationary noise,the performance of these methods become worse when the non-stationary noise is introduced.The reason why this problem happens may be that we can not exploit some prior knowledge about speech and noise in advance.In view of this problem,a wide range of methods with priori information of speech and noise have been proposed,such as codebook-driven method.For codebook-based speech enhancement methods,two codebooks of speech and noise are constructed as their priori information.The information stored in two codebooks is the trained linear predictive(LP)coefficients of speech and noise.The auto-regressive(AR)paramenters(the spectral envelope of AR and the gain of AR)are calculated online.At last,the auto-regressive(AR)Wiener filter obtained by the AR parameters is used to enhance the noisy speech.These methods work better in non-stationary noise conditions than other methods without prior knowledge.However,these methods still have shortcomings,such as,these methods model the spectral envelope as priori information,which cannot reduce the noise between harmonics.In addition,because of the big relevance between the spectral envelope of noise and the type of noise,the classification of noise is necessary in this method.Focusing on the above problems,we propose corresponding solutions in this paper.(1)With the help of the binaural cue coding(BCC)method,a speech enhancement method based on binaural cues codebook is proposed in this paper.In this method,the binaural cues of speech and noise are trained to a codebook as the prior information offline,which avoids the problem of the classification of noise.Then,the weighted codebook mapping(WCBM)algorithm is used to estimate the clean cue.At last,considering the spectral detail,we use the BCC scheme to design the gain function for enhancing the noisy speech,which can reduce the noise between the harmonics.(2)For the problem of the inaccurate estimation of clean cue,we propose a apeech enhancement method based on the binaural cues derived from deep neural network(DNN).In this method,DNN is introduced into the method to estimate the clean cue.Compared to the WCBM algorithm,DNN method can estimate the clean cue from the pre-enhanced cue derictly,which can obtain a more accurate estimation of the clean cue.In this paper,the stacked auto-encoderes often used in most of speech enhancement methods is taken into account as a type of DNN to obtain a more accurate estimation of the clean cue.In this paper,the spectrogram,PESQ(Perceptual Evaluation of Speech Quality),LSD(Log-spectral distortion)and the SSNR(Segmental signal-to-noise ratio)improvement are used to do the measurement of the enhanced speech quality.The test results show that the performance of the proposed methods outperform the conventional reference methods.