Speech enhancement by wavelet packet transform with best fitting regression line in various noise environments

Abstract

In this paper, we suggest a speech enhancement method which can be applied in various noise environments. This method uses a wavelet packet transform (WPT) and a best fitting regression line (BFRL) in order to accurately estimate parameters for the spectral subtraction method based on the time-varying gain function. It should be noted that our method does not use the statistical information of pause region detected by voice activity detector. The evaluation is performed on various environments where the noisy speech are between SNR -5 ~ 15 dB, in various noises. We compare the performance of the proposed method, with that of magnitude spectral subtraction in WPT and nonlinear magnitude spectral subtraction in WPT. We can see that the performance of the proposed method is better than that of any other methods, with regard to objective test (segmental SNR, weighted spectral slope), spectrogram analysis, and subjective one (mean opinion score). Especially, our method showed reliable result even at low SNR