This is published under the terms of CC BY-NC.

Downloaded: 687 times

Abstract

In this paper, we propose a better update algorithm for independent low-rank matrix analysis (ILRMA). ILRMA has two types of parameters, demixing vectors and non-negative matrix factorization parameters, which are estimated by minimizing the same objective function. Although many extensions of ILRMA have been proposed, the importance of the order of parameter updates in ILRMA is not investigated sufficiently. Because of the observation that iterative projection two (IP2) shows a higher performance than IP1, we propose a repeated update of demixing vectors with the source model fixed in one iteration; this approximates a simultaneous update of all demixing vectors together. We conducted music source separation experiments with more than 100 songs. The results showed that the proposed algorithm with the repeated update of demixing vectors outperforms the conventional ILRMA regarding separation performance and convergence speed.