Signal Models for Monaural Source Separation

Source separation Most real world signals contain contributions from multiple sources (e.g. cocktail party) Want to infer the original sources from the mixture Robust speech recognition Hearing aids Ron Weiss (LabROSA, Columbia University)Signal Models for Monaural Source Separation April 27, 2007 2 / 14 What makes monaural separation possible Natural sounds tend to be sparse in time and frequency- 10 % of T-F cells contain 78 % of the energy And redundant- still intelligible when 22 % of the source energy is masked “Glimpses ” of clean signal even in dense mixtures Ron Weiss (LabROSA, Columbia University)Signal Models for Monaural Source Separation April 27, 2007 3 / 14 Previous work Use constraints on the source signals to guide separation Multiple-channel case Independence constraints (e.g. ICA)- perfect separation possible Spatial constraints (e.g. beamforming) Single-channel Under-determined- more unknowns (sources) than observations Use perceptual cues similar to lower level processes in human auditory system (CASA) Segment STFT into glimpses of each source By harmonicity, common onset, etc. Sequential grouping heuristics Gaps... Inference based on prior source model