A toolbox for performance measurement in (blind) source separation

Developed with the support of the French GdR-ISIS/CNRS Workgroup ``Resources for Audio Source Separation''

by C. Févotte, R. Gribonval and E. Vincent


BSS_EVAL is a MATLAB toolbox to compute performance measures in (blind) source separation within an evaluation framework where the original sources are available as ground truth [ 1 , 2 , 3 ]. The measures are based on a decomposition of each estimated source into four contributions corresponding to the target source, interferences of unwanted sources, remaining additive noise and artifacts such as "musical noise". They are valid for all usual types of signals, such as real-valued audio or biomedical signals or complex-valued subbands of these signals.

What it can do / what it cannot do

BSS_EVAL can measure the performance of source separation algorithms, including in convolutive or time-varying settings. However, BSS_EVAL cannot separate sources: it does not implement any BSS algorithm.


Examples taken from our paper [1] are available here.

User manual

The user guide is published as a technical report (official version) both in PDF and in PostScript (minor typos fixed)

Download (distributed under the terms of the GNU GENERAL PUBLIC LICENSE)


  1. E. Vincent, C. Févotte and R. Gribonval, Performance measurement in Blind Audio Source Separation, to appear in IEEE Trans. Speech and Audio Processing.
  2. C. Févotte, R. Gribonval and E. Vincent, BSS_EVAL Toolbox User Guide -- Revision 2.0, IRISA Technical Report 1706, April 2005, Rennes, France
  3. R. Gribonval, L. Benaroya, E. Vincent and C. Févotte, Proposals for performance measurement in source separation, Proc. ICA'03, Nara, Japan, April 2003, pp 763-768

Last modified: Mon Sep 05 10:14:58 +0200 2005