Single channel sound source separation combining delay estimation and the ADRess algorithm

A method for single channel source separation is proposed in this paper, which uses estimates of the delay co-efficient of individual sources within an echoic mixture using autocorrelation, following which a "pseudo-stereo mixture" is generated, to which the ADRess algorithm can be applied. The system is evaluated in a theoretical situation, where the mixture signal to be separated consists of two individual source signals, and a delayed version of each signal. Estimates of the individual delay lengths are made and then used to create a pseudo stereo mix, where one channel consists of the original mixture signal, and the second channel consists the original mixture signal shifted by the length of the delay calculated for each source. The ADRess algorithm is then used to separate sources from the new pseudo stereo mixture.