Activation Driven Synchronized Joint Diagonalization for Underdetermined Sound Source Separation

Abstract

Blind sound source separation (BSS) is effective to improve the performance of various applications such as speech recognition. The condition of BSS can be divided into underdetermined conditions (number of microphones < number of sound sources) and overdetermined conditions (number of microphones ≥ number of sound sources). Here, we focus on Synchronized Joint Diagonalization (SJD) [6], which is a newly proposed BSS method and utilizes non-stationarity of a sound source signal. The advantage of SJD is faster separation and smaller number of parameters to be estimated. However, the application of SJD is limited to overdetermined conditions, and the performance of SJD is degraded in underdetermined conditions. In this paper, to solve these performance degradations, we propose an activation driven SJD, which uses a pre-estimated activation matrix. It is practical because activation estimation is easier than source separation. The effectiveness of the proposed method was validated by conducting BSS experiments. We confirmed that the performance of SJD can be improved in underdetermined conditions.