Noise reduction from speech signal based on wavelet transform and Kullback-Leibler divergence

A new method for noise reduction from speech signals based on the Kullback-Leibler (KL) divergence has been presented in this paper. First, the algorithm performs the wavelet packet transform to the noisy speech and decomposes it into subbands; then we apply a threshold on the noisy speech coefficients, in each subband, to obtain the enhanced speech. To determine the threshold, first the distributions of the noisy speech, estimated noise and estimated clean speech subbands are calculated; then a symmetric KL distance is calculated between the noisy speech and noise distributions. Finally a speech/noise decision is made based on the calculated distance. We conducted some tests using the TIMIT database in order to assess the performance of the proposed method and to compare it to the previous speech enhancement methods. The algorithm is evaluated using the perceptual evaluation of speech quality measure (PESQ) and the gain on input SNR. We obtain an improvement of 2db on SNR and 0.5 on PESQ-MOS score for the proposed method in comparison to the results of the previous wavelet based techniques for tested noises.