Version 2.0, October 2011. By Valentin Emiya and Emmanuel Vincent, INRIA, France.

This is joint work by Valentin Emiya (INRIA, France), Emmanuel Vincent (INRIA, France), Niklas Harlander (University of Oldenburg, Germany), Volker Hohmann (University of Oldenburg, Germany). Reduction of the computation time of the decomposition step proposed by Audionamix within the i3DMusic project (http:https://i3dmusic.audionamix.com).

The PEASS Software provides a set of perceptually motivated objective measures for the evaluation of audio source separation.

Similarly to BSS Eval, the distortion signal is decomposed into three components: target distortion, interference, artifacts. These components are then used to compute four quality scores, namely OPS (Overall Perceptual Score), TPS (Target-related Perceptual Score), IPS (Interference-related Perceptual Score), APS (Artifact-related Perceptual Score). These scores better correlate with human assessments than the SDR/ISR/SIR/SAR measures of BSS Eval.

This method can be applied either to the source signals or to their spatial images, whatever the number of channels.

• Unzip the file into a new directory.

• Download the third-party gammatone filterbank and haircell modeling software and unzip it in the above directory. Please pay attention to the specific user license agreement governing this third-party software. Under Bash, do the following steps:

  $ wget http:https://medi.uni-oldenburg.de/download/demo/adaption-loops/adapt_loop.zip
  $ unzip adapt_loop.zip
  $ wget http:https://medi.uni-oldenburg.de/download/demo/gammatone-filterbank/gammatone_filterbank-1.1.zip
  $ unzip gammatone_filterbank-1.1.zip

• Compile the MEX files by running compile under Matlab (this is optional but leads to much faster computation).

The main function is PEASS_ObjectiveMeasure.m. For an example of use, see example.m. The distortion components are stored as .wav files in the example/ directory.

The code can be used on any platform where Matlab is installed.

Please report any bug or comment to [email protected] and [email protected]. So far, some technical limitations have been noticed solved (but maybe not optimally yet):

• out of memory/large sound materials: when sounds are long, sampling frequency is high and/or sources are numerous, an "out of memory" issue may be raised. In this case, increase the option.segmentationFactor integer value to have the sounds segmented first, then decomposed and finally merged along the full time scale. This is due to the gammatone implementation and the current solution may be improved in the future.

When using this software, the following papers must be referred to:

• Valentin Emiya, Emmanuel Vincent, Niklas Harlander and Volker Hohmann, Subjective and objective quality assessment of audio source separation, IEEE Transactions on Audio, Speech and Language Processing, 19(7):2046-2057, 2011.
• Emmanuel Vincent, Improved perceptual metrics for the evaluation of audio source separation, 10th Int. Conf. on Latent Variable Analysis and Signal Separation (LVA/ICA), pp.430-437, 2012.

The toolbox uses an implementation of the gammatone filterbank presented in

• V. Hohmann, Frequency analysis and synthesis using a Gammatone filterbank, Acustica/Acta Acustica, 88(3):433-442, 2002
• T. Herzke and V. Hohmann, Improved numerical methods for gammatone filterbank analysis and synthesis, Acustica/Acta Acustica, 93(3):498-500, 2007

The toolbox uses the hair cell model and the PEMO-Q metrics described in

• T. Dau, B. Kollmeier and A. Kohlrausch, Modeling auditory processing of amplitude modulation: I. Modulation Detection and masking with narrowband carriers, J. Acoust. Soc. Am., 102(5):2892-2905, 1997
• R. Huber and B. Kollmeier, PEMO-Q -- A New Method for Objective Audio Quality Assessment Using a Model of Auditory Perception, IEEE Trans. on Audio, Speech, and Language Processing, 14(6):1902-1911, 2006

Copyright 2010-2011 Valentin Emiya and Emmanuel Vincent (INRIA).

The code in the current directory is distributed under the terms of the GNU Public License version 3 (http:https://www.gnu.org/licenses/gpl.txt).

Version 2.0, October 2011:

• changed some parameters of the decomposition and of PEMO-Q
• changed the training procedure

Version 1.1, September 2011:

• replaced the PEMO-Q software by a Matlab/MEX implementation (audioQualityFeatures.m).
• forced the subjective scores to 100 for hidden references in the training stage

Version 1.0.1, September 2011:

• added an error message if signal sizes are not correct (extractDistortionComponents.m).
• improved the processing of multichannel signals (audioQualityFeatures.m).

Version 1.0, May 12th, 2010:

• first release.