Blind Source Separation/ Localization/ Spatialization

    University of Bordeaux 1 (science and technology)

  This page includes many listenings examples issued from our complete computational system for Auditory Scene Analysis (CASA-EM). The system localize, separate and spatialize an arbitrary number of audio sources given only binaural signals. We assume that the short-time spectra of the sources in the mixtures do not overlap significantly, a hypothesis, called Windowed-Disjoint-Orthogonality (WDO). The source separation estimate the filtered sources via clustering of the interaural cues. The two spatialization algorithms (SSPA, MHRTF) simulate a source as originating from any direction of arrival angle (DOA).
The examples enables each listener to assess subjectively the performance of spatialization and separation components in the horizonthal plane. 
  The complete system will be fully describe in a next report available here. The report will contain all theoretical aspects, pratical and implementation recommandations. Simulations results. Possible research orientations to further improve the whole system effectiveness.


Degenarate Demixing with CASA-EM



Comparison of CASA-EM demixing with other BSS methods (DUET,...)

Mixtures : xL.wav | xR.wav

Demixtures :

Source Spatialization with CASA-EM : Two techniques --> SSPA and MHRTF

"Spatializures" :