Research On Speech Enhancement Algorithm Based On Prior Signal-to-noise Ratio Estimation

With the development of artificial intelligence,as one of the main means of information transmission,speech interaction becomes more and more important.And we live in a noisy word,it is not avoid that the quality of the picking voice has seriously affected by the external noise.In order to solve this problem,we can obtain clean speech from the noisy speech by selecting the appropriate speech enhancement algorithm,which is carry out in the front-end to improve the performance of the system.Speech enhancement has been widely used in human-computer interaction,which plays an important role in voice picking and recognition,speech coding and military communication.Based on the demand of realistic scene,this thesis focuses on the single channel speech enhancement method.According to the deficiency of the a priori SNR estimators,some new methods for the a priori SNR estimator are proposed from the perspective of harmonic reconstruction and human ear masking effect.The feasibility of the proposed algorithm is analyzed from both theoretical and experimental analysis.The main work and innovation of this thesis can be reflected in the following aspects:1)The a priori SNR estimator based on harmonic regeneration is proposed in this thesis.The a priori SNR estimator of single-channel speech enhancement can effectively remove the noise when the SNR is high,but the speech distortion is more serious when the SNR is low.In the algorithm,the enhanced signal is processed by the secondary spectrum,which can enhance the periodicity of the speech signal;and then the harmonic regeneration is carried out,thus the harmonic component is enhanced.Experimental results show that the proposed algorithm can effectively enhance the harmonic component of the speech signal in low SNR,and has less speech distortion than the traditional SNR estimators.2)A single channel speech enhancement algorithm for noisy speech signals is proposed,which is based on reassigned spectrogram and masking properties of the human auditory system.In order to solve the residual background noise and the musical noise resulted by the a priori SNR estimator,according to the strong correlation of speech harmonics,we propose to use the correlation of adjacent frames and that of harmonics of the reassigned spectrogram to control the forgetting factor of the two-step-decision-directed method,which can suppress the non-speech components better.Then,the estimated speech spectra is employed to compute the noise masking threshold of a perceptual gain factor by the masking properties of the human ear.Experimental results show that the proposed algorithm can improve the intelligibility of speech signals with the same amount of noise reduction.3)Speech enhancement algorithm of binary mask estimation based on priori SNR constraints is proposed.We first analyzed the impact of estimation of the a priori SNR on the noise spectrum estimation function with low SNR.And the MMSE rule is used to modify the a priori SNR by secondary processing,which can get the estimated noise power spectrum and gain function better;and the estimate is employed to retain noise over-estimated T-F units while discarding noise under-estimated T-F units.Experiments results show that the proposed algorithm can improve the intelligibility of speech signals with low SNR.