Wavelet-based underdetermined blind source separation of speech mixtures

Blind Source Separation (BSS) is an approach to estimate original source signals only by observed signals through multiple sensors without any prior knowledge about their mixture process. The BSS is currently expected to be applied to such a broad field as preprocessing of speech recognition, analysis of bio-signals, etc. In this paper, we focus on underdetermined BSS (i.e., BSS with more sources than sensors). In underdetermined BSS, there are many estimation algorithms based on sparse representation of source signals. The accuracy of decomposition of source signals depends strongly on sparsity hypothesis. For this purpose, it is effective to preprocess observed signals by using a linear transformation with respect to the base so as to make it more sparse representation of signals. We now propose an application of time-frequency analysis to preprocessing of sound source signals using the RI-Spline wavelet transform. The effectiveness of the proposed approach is demonstrated through a computer simulation using real data.