Optimization of the asymptotic performance of time-domain convolutive source separation algorithms

摘要

In this paper, we investigate the self-adaptive source separation problem for convolutively mixed signals. The proposed approach uses a recurrent structure adapted by a generic rule involving arbitrary separating functions, which is derived from a neural approach. The expression of the asymptotic error variance achieved by this rule is first determined (for strictly causal mixtures). This enables us to derive the separating functions that minimize this error variance. They are shown to be only related to the probability density functions of the sources. Simulations are performed in various conditions, ranging from artificial mixtures of synthetic sources to real mixtures of audio signals. They show that the proposed approach yields much better performance than classical rules.