Research Of Effient Speech Enhancement And Voice Activity Detection

The speech processing system is usually interfered by various kinds ofnoise in the practical environment. For the diversity of noise power and model,the performance of speech coding or recognition has large loss in the noisycondition. At the same time, the computation time and memery are limited inthe mobile communication. So using speech enhancement and voice activitydetection as front-end processing is a useful way to improve the performancein the noise condition.In this dissertation, the algorithm is research from three aspects,including noise estimation, voice activity detection and speech enhancement.A low complexity and stable system for processing noisy speech is proposed.Firstly, because the algorithm of minimum statistics noise estimation ishigh complexity and can not tracking very non-stationary noise rapidly, analgorithm of noise estimation with adaptive smooth factor is proposed. Thesmooth factor is estimated base on soft decision of speech present probability.The speech present probability is estimated base on posterior SNR (Signal toNoise Ratio), then the smooth factor is estimated by using the soft informationof speech present probability, and the noise spectrum is estimated at last. Theexperiment demonstrates that the NMSE (Normalized Mean Squared Error) ofthe proposed algorithm is lower than MS (Minimum Statistics) method inwhite and babble noisy condition. Also the algorithm has lower complexitybecause no need for searching minimum.Secondly, because the tranditional voice activity detection can not workwell in non-stationary noisy condition, an algorithm of VAD (Voice ActivityDetection) based on SNR and whitening filtered entropy is proposed. By usingthe estimated noise spectrum, the whitening filter is applied to get the entropy.Also with the help of SNR, the VAD result is decided. The threshold isadjusted on SNR and the entropy. The experiment demonstrates that the error rate of the proposed VAD algorithm is lower than that of G.729B and AMRVAD1.Finally, because the algorithm of speech enhancement base ofLSA-MMSE has high complexity, a new gain function base on parametercontrol is proposed. The gain is mainly controlled by prior SNR andcompensed by posterior SNR. The experiment shows that the proposedalgorithm gets higher PESQ scores than the LSA-MMSE method.