Speech Enhancement Technique Research In Low SNR Conditions Based On Short-Time Spectrum Estimation

Speech enhancement techinique aims to improve the speech quality, which is employedmore and more often in new applications of speech recognition and voice communication. As theSignal-to-Noise Ratio (SNR) degrades, people need speech enhancement more greatly. But inlow SNR conditions, the compromise among speech distortion and original residual noise is hardto determine, and the noise-robust methods are few. This paper aims at the speech enhancementalgorithms of spectral enhancement in low SNR conditions based on multi-type of noise, and themain work is generalized as follows:Firstly, the spectral enhancement algorithms were analyzed. Gain function、a prior SNR、speech present probability were analyzed separately, and various speech quality assessmentmethods were discussed. A testing platform was promosed, based on which some algorithms suitfor low SNR conditions were compared, the conclusion was given that the OM-LSA algorithmusing speech present probability performed better than others in low SNR conditions, and thenoise estimation was crucial.Secondly, the noise estimation algorithms were researched. Large delay of the minimatracking algorithms which are suit for the low SNR conditions were analyzed, and the improvedminima controlled recursive averaging algorithm was improved in the article. The classicalalgorithms were usually limited to the use of length-fxed search window, and in the proposedalgorithm, the speech present probability was calculated iteratively to determine the length ofsearch window. Meanwhile, the speech present probability was adjudged again and the safety netwas introduced. The simulated experiments indicated that the algorithm was able to estimatenoise more accurately by reducing the delay of the noise minima tracking.Thirdly, we proposed a modified method in order to reduce the influence of addictive noiseon the phase spectrum of the estimation of clean speech. We contrasted the estimated phasespectrum with that of clean speech for the purpose of showing the influence of addictive noise,and an algorithm refined speech phase estimation to boost the performance of speechenhancement algorithms in low SNR conditions was proposed. The compensated short timephase spectrum was computed as follows: The noise spectral magnitude estimator was derivedfrom the speech magnitude estimator for the symmetry of the statistical model; then we obtainedimprove the phase spectrum estimation function in each frequency bin; In the end, a modifiedenhancement algorithm was proposed based on OM-LSA algorithm. The simulated experimentsindicated that significant improvement could be achieved. For the speech enhancement in lowSNR environment, the algorithm obtained an obvious improvement in speech quality, with littlemusical noise retained.At last, in order to get better enhancement result, an algorithm to reduce speech distortionwas proposed with the improved noise estimation and phase spectrum estimation which werebrought forward. Drawback of OM-LSA algorithm was analyzed, and MMSE spectral magnitudegain function was use to improve OM-LSA algorithm to reduce speech distortion. Then a modified estimator for the a priori SNR was proposed. The testing results under multi-type ofnoise showed that the proposed approach worked well, which not only less residual noise butalso less speech distortion than OM-LSA algorithm.