Steven Rennie, Parham Aarabi, Trausti Kristjansson, Brendan J. Frey, Kannan Achan
A variational inference algorithm for robust speech separa-tion, capable of recovering the underlying speech sourceseven in the case of more sources than microphone obser-vations, is presented. The algorithm is based upon an gen-erative probabilistic model that fuses time-delay of arrival(TDOA) information with prior information about the speak-ers and application, to produce an optimal estimate of theunderlying speech sources. Simulation results are presentedfor the case of two, three and four underlying sources andtwo microphones observations corrupted by noise. The re-sulting SNR gains (24dB with two sources, 15dB with threesources, and 9dB with four sources) are significantly higherthan previous speech separation techniques.