A Wavelet Approach to Convolutive Blind Separation of Non-stationary Sound Sources

Authors:	Takada Kiyotaka, The University of Electro-Communications, Japan Nakano Kazushi, The University of Electro-Communications, Japan Watai Hirokazu, The University of Electro-Communications, Japan
Topic:	1.1 Modelling, Identification & Signal Processing
Session:	Filtering and Estimation
Keywords:	blind source separation, convolutive mixing process, wavelet transform, Gabor function, InfoMax method, cross-correlation method

Abstract

This paper considers a wavelet-based approach to the blind separation of the signals (BSS) including reflections such as sound. The BSS for the instantaneous mixture case is achieved by maximizing the mutual information between input and output signals or by zeroing the cross-correlations of the separated signals. For making these methods applicable to the convolutive mixture case, it is important to transform signals in the time domain into those in the time-frequency domain by the wavelet transforms with the Gabor function.　Simulation results demonstrate the effectiveness of our wavelet-based approach.