Normalized observation vector clustering approach for sparse source separation

Shoko Araki, Hiroshi Sawada, Ryo Mukai, Shoji Makino

Abstract: This paper presents a new method for the blind separation of sparse sources whose number N can exceed the number of sensors M. Recently, sparseness based blind separation has been actively studied. However, most methods utilize a linear sensor array (or only two sensors), and therefore have certain limitations; e.g., they cannot be applied to symmetrically positioned sources. To allow the use of more than two sensors that can be arranged in a non-linear/non-uniform way, we propose a new method that includes the normalization and clustering of the observation vectors. We report promising results for the speech separation of 3-dimensionally distributed five sources with a non-linear/non-uniform array of four sensors in a room (RT <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">60</sub> = 120 ms).