Lightweight Online Separation of the Sound Source of Interest through BLSTM-Based Binary Masking

Abstract

Online audio source separation has beenan important part of auditory scene analysis and robotaudition. The main type of technique to carry thisout, because of its online capabilities, has been spatialfiltering (or beamforming), where it is assumed that thelocation (mainly, the direction of arrival; DOA) of thesource of interest (SOI) is known. However, thesetechniques suffer from considerable interference leakagein the final result. In this paper, we propose a two steptechnique: 1) a phase-based beamformer that provides,in addition to the estimation of the SOI, an estimationof the cumulative environmental interference; and 2) aBLSTM-based TF binary masking stage that calculatesa binary mask that aims to separate the SOI from thecumulative environmental interference. In our tests, thistechnique provides a signal-to-interference ratio (SIR)above 20 dB with simulated data. Because of thenature of the beamformer outputs, the label permutationproblem is handled from the beginning. This makes theproposed solution a lightweight alternative that requiresconsiderably less computational resources (almost anorder of magnitude) compared to current deep-learningbased techniques, while providing a comparable SIRperformance.