A New Speech Denoising Method Based on WPD-ICA Feature Extraction

Abstract

Independent Component Analysis (ICA) feature extraction is an efficient sparse coding method for noise suppression. However, single channel signal can not be directly applied in ICA feature extraction. In this paper, we propose a new method using wavelet packet decomposition (WPD) as preprocessing for single channel data. Wavelet packet coefficients (WPCs) provide multi-channel data as input data to learn ICA basis vectors. Furthermore we project input data onto the basis vectors to get sparser and independent coefficients. Appropriate nonlinear shrinkage function is used onto the components of sparse coefficients so as to reduce noise. The proposed approach is very efficient with respect to signal recovery from noisy data because not only the projection coefficients are sparser based on WPCs but both the features and the shrinkage function are directly estimated from the observed data. The experimental results have shown that it has excellent performance on signal to noise ratio (SNR) enhancement compared with other filtering methods.