Combined discrete wavelet transform and wavelet packet decomposition for speech enhancement

A new wavelet domain algorithm for speech enhancement is presented. The main idea of the proposed algorithm can be described as follows: noisy speech is first split into low frequency band signal and high frequency band signal via one scale of DWT. Then, three scales of DWT and three scales of WPD are applied to the low frequency signal and the high frequency signal respectively. The denoised speech is reconstructed finally. To prevent the degradation of unvoiced sounds during noise reduction, this algorithm divides noisy speech samples into unvoiced segments and voiced segments and then applies multi-threshold to them. Compared with DWT and WPD, the new algorithm can preserve more unvoiced sounds and provide better performance as shown by computer simulation results.