Research On Monaural Speech Enhancement Based On DNN And Phase Spectrum Compensation

Speech enhancement as the front-end processing module of speech recognition,aimed at curbing the speech with noise in composition,improve the whole quality of speech enhanced sense of listening to and play an important role in the field of human-computer interaction.At present,most speech enhancement algorithms are supervisory algorithms,which do not need to know prior information about speech and noise,and directly use the noisy speech to get the target speech.However,the supervised algorithm also has some problems.The noise in real life is mostly non-stationary noise.The single mapping relationship obtained by the original learning is not suitable for all noises,resulting in inaccurate noise estimation and unsatisfactory noise suppression effect,thus affecting the overall quality of speech.In addition,phase information is often ignored by current speech enhancement algorithms.However,studies have found that if the phase information of the pure speech is introduced into the enhancement algorithm,the intelligibility of the speech will be significantly improved,so the phase spectrum of the pure speech should be estimated before signal reconstruction.In view of the current problems of speech enhancement algorithms,this thesis has done research work is as follows:(1)In view of the problem of poor noise suppression effect of current speech enhancement algorithms,a time-frequency mask algorithm based on deep neural network(DNN)combined with convex optimization is studied for monaural speech enhancement.Firstly,the algorithm extracted with noise energy spectrum as input within DNN characteristics of speech;Secondly,the inter-channel correlation factor(ICC Factor)between noise and noisy speech is used as the training target of DNN;Then,reconstruct the objective function of convex optimization through the correlation factor obtained from the DNN model;Finally,new hybrid conjugate gradient method based on DNN combined with convex optimization,is used to update the initial mask.The final updated mask is used to synthetic the enhanced speech.It can be seen from the simulation experiment that under low signal-to-noise ratio conditions with different noises,the log spectral distance(LSD),perceptual evaluation of speech quality(PESQ),short-time objective telligibility(STOI)and segmental signal to noise ratio(seg SNR) indicators of enhanced speech are better than before the improvement,and noise suppression effect and the overall speech quality have been improved.(2)Aiming at the traditional phase spectrum compensation(PSC)speech enhancement algorithm that uses a fixed phase compensation factor,and the performance of the algorithm is closely related to the accuracy of noise estimation.A sparsity-based phase spectrum compensation(SPSC)speech enhancement algorithm is studied.First,obtain the noise amplitude spectrum and the enhanced speech amplitude spectrum through the noise estimation algorithm and the speech enhancement algorithm based on the amplitude spectrum respectively;Secondly,calculate the spectral time sparsity of the current time-frequency unit through the local signal-to-noise ratio(SNR)between the noise and the target speech amplitude spectrum;Then,the SPSC function is obtained by combining the spectral time sparsity and the phase compensation factor improved by the sigmoid function;Finally,use the SPSC function to compensate the spectral components in the phase spectrum,and the enhanced speech of this thesis is obtained through inverse short-time fourier transform(ISTFT).The simulation results show that under four different background noises with low signal-to-noise ratio,the enhanced speech LSD,PESQ and seg SNR indicators obtained by the new phase spectrum compensation algorithm are better.It shows that under low signal-to-noise ratio,the new algorithm has a significant denoising effect and can effectively restore the speech components in noisy speech,enhancing the quality of speech and improving the sense of hearing.