A wavelet fusion method for speech enhancement

In order to integrate the characteristics of different noise reduction algorithms, a wavelet fusion method for speech enhancement is proposed in this paper. In the proposed method, the noisy speech is first decomposed into several sub-bands by wavelet packet analysis, and then enhanced by the statistical model based method and wavelet thresholding method, respectively. The output of each sub-band is obtained under the fusion rule based on the cross-correlation and the a priori SNRs of the two enhanced coefficient sets. Finally, the enhanced coefficients are transformed back to time domain to get the enhanced speech. The performance of the proposed method is evaluated under ITU-T G.160. The test result shows that, comparing with the reference algorithms, the proposed method produces enhanced speech with better objective quality, while the amount of noise reduction and SNR improvement remains at a high level. The impact on the speech level is kept in an acceptable range at the same time.