The corresponding participants are:
* Antoine Liutkus,  Telecom ParisTech, CNRS LTCI, Paris, FRANCE
* Zafar Rafii, Northwestern University EECS Department Evanston, IL, USA
* Roland Badeau; Telecom ParisTech, CNRS LTCI, Paris, FRANCE
* Bryan Pardo, Northwestern University EECS Department Evanston, IL, USA
* Gaėl Richard, Telecom ParisTech, CNRS LTCI, Paris, FRANCE

(Note that Zafar and Bryan are involved in the separation of the vocals
only. Antoine, Roland and Gaėl are involved in all signals.)


The algorithm is based on a specialized technique per signal, i.e:
vocals, bass, drums and other. The vocals are extracted using the
adaptive REPET algorithm, which will be submitted to ICASSP 2012. The
drums are extracted using the technique discussed in:

- Gaussian Processes for Underdetermined Source Separation, IEEE Trans.on Sig. Proc. 59(7), July 2011

The bass is extracted by first computing its time-varying fundamental
frequency, then by designing a soft Time-Frequency mask to remove the
corresponding harmonics.

The whole process has been designed to be computationally rather
efficent: all separated files for one mixture are computed in about a
minute, on one 3GhZ core with 8gb RAM.