Underdetermined source separation using a sparse STFT framework and weighted laplacian directional modelling

TL;DR

This paper introduces a novel sparse STFT framework with weighted Laplacian directional modeling for underdetermined audio source separation, improving accuracy over previous methods and competing well with current state-of-the-art techniques.

Contribution

It develops a weighted mixture of directional Laplacian densities in a sparse STFT domain for better underdetermined source separation.

Findings

Improved separation results over previous methods

Favorable comparison with state-of-the-art techniques

Enhanced performance in sparse STFT domain

Abstract

The instantaneous underdetermined audio source separation problem of K-sensors, L-sources mixing scenario (where K < L) has been addressed by many different approaches, provided the sources remain quite distinct in the virtual positioning space spanned by the sensors. This problem can be tackled as a directional clustering problem along the source position angles in the mixture. The use of Generalised Directional Laplacian Densities (DLD) in the MDCT domain for underdetermined source separation has been proposed before. Here, we derive weighted mixtures of DLDs in a sparser representation of the data in the STFT domain to perform separation. The proposed approach yields improved results compared to our previous offering and compares favourably with the state-of-the-art.