Relaxing the WDO Assumption in Blind Extraction of Speakers from Speech Mixtures

Abstract

The time-frequency masking approach in blind speech extraction consists of two main steps: feature clustering in a space spanned over delay-time and attenuation rate, and spectrogram masking in order to reconstruct the sources. Usually a binary mask is generated under the strong W-disjoint orthogonal (WDO) assumption (disjoint orthogonal representations in the frequency domain). In practice, this assumption is most often violated leading to weak quality of reconstructed sources. In this paper we propose the WDO to be relaxed by allowing some frequency bins to be shared by both sources. As we detect instantaneous fundamental frequencies the mask creation is supported by exploring a harmonic structure of speech. The proposed method is proved to be effective and reliable in experiments with both simulated and real acquired mixtures.