Speech enhancement using a wavelet thresholding method based on symmetric Kullback-Leibler divergence

Performance of wavelet thresholding methods for speech enhancement strongly depends on estimating an exact threshold value in the wavelet sub-bands. In this paper, we propose a new method for more exact estimation of the threshold value. Our proposed threshold value is firstly obtained based on the symmetric Kullback-Leibler divergence between the probability distributions of noisy speech and noise wavelet coefficients. In the next step, we improved this value using the segmental Signal-to-Noise Ratio (SNR). We used some TIMIT utterances to assess the performance of the proposed threshold. The algorithm is evaluated using the Perceptual Evaluation of Speech Quality (PESQ) score and the SNR improvement in ideal and real modes. In ideal and real modes, on average, we obtain respectively 2.25 dB and 1 dB SNR improvement and a PESQ score increase up to 1.1, 0.75 compared with the conventional wavelet thresholding approaches. In comparison to the adaptive thresholding approach, on average in ideal and real modes, we obtain respectively 1.6 dB and 0.9 dB SNR improvement. The PESQ value of the adaptive thresholding method, in the real and ideal modes, is 0.25 higher and 0.5 lower than that of our proposed method, respectively.