Fast Monaural Separation of Speech

Niels Henrik Pontoppidan, Mads Dyrholm

AbstractWe have investigated the possibility of separating signals from a
single mixture of sources. This problem is termed the Monaural
Separation Problem.

Lars Kai Hansen has argued that this problem is topological tougher than
problems with multiple recordings.

Roweis has shown that inference from a Factorial Hidden Markov Model, with non-stationary assumptions on the source autocorrelations
modelled through the Factorial Hidden Markov Model, leads to
separation in the monaural case.

By extending Hansens work we find that Roweis' assumptions are necessary for monaural speech separation.

Furthermore we develop a Factorial hierarchical vector quantizer
yielding a significant decrease in complexity of inference.
KeywordsMonaural, Independent Component Analysis, Blind Source Separation, Factorial Hidden Markov Model, Non-stationarity
TypeConference paper [With referee]
ConferenceAES 23rd International Conference, Signal Processing in Audio Recording and Reproduction.
EditorsPer Rubak
Year2003    Month May
PublisherAudio Engineering Society, Inc.
Address60 East 42nd Street, Room 2520, New York, New York, 10165 - 2520 USA
Electronic version(s)[pdf]
BibTeX data [bibtex]
IMM Group(s)Intelligent Signal Processing