Novel Methods to Improve Blind Audio Source Separation of Anechoic Mixtures Online Using DUET

Abstract

newlineSeparating individual sound sources from a complex acoustic mixture known as audio source separation has become very popular in the recent decades. Blind Source Separation (BSS) methods try to emulate the sophistication of the human auditory system which is capable of separating an unknown number of speakers from the cacophony of a crowded room. It has always been a challenging problem to solve and many BSS algorithms have been proposed as a solution to this. newlineThe objective of this research is to develop novel approaches to improve audio source separation in Degenerate Unmixing Estimation Technique (DUET), to let people access, manipulate and enjoy audio more easily. This thesis examines some of the problems faced by DUET algorithm in real world environment. newlineThe DUET is a BSS method that can separate arbitrary number of unknown sources from two microphone recordings. DUET builds a 2 dimensional histogram from the amplitude ratio and phase diand#64256;erence between channels where the peak indicates a source, with peak locations corresponding to mixing parameters associated with that source. DUET partitions the TF representation of the mixtures by allocating each bin to the source with the closest mixing parameters, assuming that the Time Frequency (TF) bins do not overlap too much, which is commonly validated by speech mixtures. However, when the TF bins of the source start overlapping too much, as in real world scenarios, DUET cannot perform source separation eand#64256;ectively. newline newline