The Research Of Speech Enhancement Algorithms Based On Wavelet Transform

As an important noise suppression tool, the main goal of speech enhancement isto restore the original speech signal from the noisy speech as much as possible. Mostof traditional speech enhancement approaches are based on short time Fouriertransform. As an emerging time-frequency analysis method, wavelet transformovercomes the disadvantage of short time Fourier transform—single resolution,providing high time resolution in the high frequency part and high frequencyresolution in the low frequency part, which makes it very advantageous for analyzingnon-stationary, time-varying signals, such as speech signals. Hence, the study ofspeech enhancement algorithms based on wavelet transform is of great significance.The research of speech enhancement algorithms based on wavelet transform inthis paper mainly includes two aspects: the combination of wavelet transform withtraditional speech enhancement algorithms and the study of adaptive waveletshrinkage method.The major drawbacks of speech enhancement in wavelet package domain aremuch signal distortion (because of down sampling at each level of decomposition) andhigh computation load. In this paper, we proposed a new speech enhancement system.This system integrates undecimated perceptual wavelet packet decomposition,OM-LSA (Optimally-modified Log-Spectral Amplitude) and IMCRA (ImprovedMinima Controlled Recursive Averaging) estimation. Experiments conducted invarious noise environments show that, compared with the OM-LSA algorithm, thenew speech enhancement system greatly suppresses the background noise and getsbetter speech intelligibility and perceptual quality without causing considerable signaldistortion and “musical noise”.For the traditional wavelet thresholding method generally has over thresholdingphenomena, whether the threshold is global or node, this paper studied the adaptivethreshold method based on teager energy operator. According to the distributioncharacteristic of speech and noise, a new speech enhancement method was proposed.Experiments’ simulation under different noise environments prove that, comparedwith the original algorithm, the new speech enhancement method not only improvesthe SNR (signal to noise ratio) and perceptual quality but also reduces thecomputation load, making it easier to realize. It is also found that the new speech enhancement algorithm is very suitable for car noise environments.