Permutation-free clustering of relative transfer function features for blind source separation

Abstract

This paper describes an application of relative transfer functions (RTFs) to underdetermined blind source separation (BSS). A clustering-based BSS approach has the advantage that it can even deal with the underdetermined case, where the sources outnumber the microphones. Among others, clustering of a normalized observation vector (NOV) has proven effective for BSS even under reverberation. We here point out that the NOV gives information about RTFs of the dominant source, and hence call it the RTF features. Most of the previous BSS methods are limited in that they undergo significant performance degradation when the number of sources is not known precisely. This paper introduces our recently developed method for joint BSS and source counting based on permutation-free clustering of the RTF features. We demonstrate the effectiveness of the method in experiments with reverberant mixtures of an unknown number of sources with a reverberation time of up to 440 ms.