Research On Speech Enhancement Algorithm Based On Improved Spectral Subtraction

Speech is an important medium for human beings to exchange information.However,in the actual speech communication process,various environmental noises and equipment noise always interfere with the transmission of information and reduce the communication quality.Therefore,in order to effectively use the speech signal,it is necessary to extract useful information from the noisy speech as much as possible,and speech enhancement is an important method to solve this case.In this thesis,some existing typical speech enhancement algorithms are analyzed and discussed,and two new algorithms are proposed on this basis.The following is the main work of this thesis:1.An improved spectral subtraction algorithm based on the masking effect of human ear and Bayesian estimation is proposed for speech enhancement.This algorithm uses IMCRA(the Improved Minima Controlled Recursive Averaging)algorithm to estimate the background noise.The algorithm performs two spectral subtractions on noisy speech,the first one is common noise power spectrum,the second one is spectral subtraction of spectral subtraction which adjusts the subtraction factors according to the masking effect of human ear.Also,the gain function of the second one is improved in this algorithm.Between these two spectral subtractions the noisy signal is enhanced with a Bayesian estimation based on a weighted likelihood ratio(WLR)distortion measure.The experimental results show that the improved algorithm can ensure the intelligibility of enhanced speech and achieve better speech enhancement while suppressing background noise and residual noise.2.An enhancement algorithm based on MMSE estimation and improved multi-band spectral subtraction is proposed.The algorithm combines logarithmic MMSE estimation with preprocessing of multi-band spectral subtraction,optimizes the smoothness of the initial signal spectrum,and uses the improved over-subtraction factor and subtraction factor function for speech enhancement.In the proposed algorithm,the minimum statistics(MS)algorithm is used to estimate the noise of logarithmic MMSE estimation,while the improved multi-band spectral subtraction algorithm combines the recursive function of voice activity detection(VAD)to estimate the noise.The simulation results show that the proposed algorithm has excellent performance,especially in the case of low signal-to-noise ratio,better suppressing noise while preserving voice information,and improving the quality of enhanced speech.