AMPLITUDE MODULATION DECORRELATION FOR CONVOLUTIVE BLIND SOURCE SEPARATION

Jörn Anemüller and Birger Kollmeier
ane@unioldenburg.de, http://medi.unioldenburg.de/members/ane

The problem of blind separation of a convolutive mixture of speech signals is considered. Signal separation is performed in the frequency domain. Based on observations from amplitude spectrograms of speech signals, the notion of amplitude modulation corre- lation (`AMCor') across dierent frequency channels is in- troduced. From the corresponding principle of amplitude modulation decorrelation, a novel costfunction and an al- gorithm for convolutive blind source separation are derived. The algorithms' main features are discussed. Successful sep- aration of synthetic data and of realroom recordings of speech is performed. The results of the latter are compared to the performance of previous algorithms on the same data. Audio examples are available from the authors' web page [2].