Authors:
Ricard Marxer and Jordi Janer


INFORMATION
----------------------------
This submission combines different algorithms that rely on model-based 
inference, taking into account the musical signal's harmonic structure. 
All methods require no human intervention.

We apply two different algorithms: one for voice separation and another 
for drums separation. Therefore a chain process of the two algorithms 
allow us to separate: "vocals", "drums" and "other instruments". 
Depending on the audio example, we will be able to separate all required 
instruments. In other cases only three separated sources will be provided.

For vocals separation, we used a Tikhonov regularization signal 
decomposition method [1]. The factorization sacrifices the 
non-negativity condition on the solution in exchange for a lighter 
computational load. The predominant pitch detection does not perform 
vocal/non-vocal classification. Average running time 32 sec/excerpt (on 
a 3.4GHz CPU).

For drums separation, the presented algorithm combines a harmonic-based 
decomposition using a Non-negative Matrix Factorization (NMF) algorithm, 
with the transient analysis of spectral peaks from a single audio frame. [2]


[1] 'A TIKHONOV REGULARIZATION METHOD FOR SPECTRUM DECOMPOSITION IN LOW 
LATENCY AUDIO SOURCE SEPARATION' Marxer, R. and Janer J. (ICASSP 2012 
submitted)
[2] 'COMBINING A HARMONIC-BASED NMF DECOMPOSITION WITH TRANSIENT 
ANALYSIS FOR INSTANTANEOUS PERCUSSION SEPARATION', Janer J., Marxer, R. 
and Arimoto, K.  (ICASSP 2012 submitted)

--