Study On Speech Enahcement In The Presence Of Noise Based On Wavelet Method

Speech enhancement is one of the key technique of speech recognition fornoisy conditions and noisy speech recognition is the active research areas forspeech recognition. At present, three general techniques are used for noisyspeech recognition: model compensation, speech enhancement and high robustcharacteristics extraction. The research in this thesis concerned about theproblem of speech enhancement in noisy environments.In this dissertation, the speech enhancement technique is researched andwavelet threshold method for de-noising and enhancement is focused on.Firstly we introduce the uniform speech enhancement algorithms and the theoryof wavelet threshold used for signal de-noising. Then put forward acompromise threshold which gains a better performance than hard thresholdand soft threshold. On this base, the result for de-noising and enhancementusing the threshold function under five different threshold conditions isproposed, and the de-noising results under choosing different thresholdfunction condition are compared.The collection of sound signal is based on LabVIEW and arithmeticemulation is based on Matlab. The emulation result indicates that through thecompare of MSE and enhancement of Signal-to-Noise, the re-constructedsignal using the compromise threshold function for de-noising andenhancement of sound signal mentioned in this paper is more smooth amongthe three threshold functions, and the MSE of de-noised signal based onNeyman-Pearson is the lowest among the five threshold rules.The distortion problem brought from sound enhancement is discussed inthis paper. The result of distortion include residuals approximately representedadditive and multiplicative noise aroused by spectrum distortion. The distortioncomes from multiplicative noise is compensated using CMN in characteristicspace, and the distortion results from additive residuals is compensated usingPMC in model space. Then the recognition rate proved the recognitionefficiency.