[ file USS-EXAMPLE/README ]

Example of the Unsupervised Spectral Subtraction scheme proposed in:

"Unsupervised Spectral Subtraction",
by G. Lathoud, M. Magimai.-Doss, B. Mesot and H. Bourlard,
to appear in Proceedings of ASRU 2005.

----------
CONTENTS

[ USS-EXAMPLE.zip ]

An archive file containing this entire directory and its subdirectories.

[ DATA ]
Directory containing various WAV files,
the "av163*wav" files come from the AV16.3 corpus (http://mmm.idiap.ch/Lathoud/av16.3_v6/AV163.pdf),
the "*NU-1028*wav" files come from the Numbers 95 corpus, with some NOISEX Factory or Lynx additive noise.

[ README ]
This file.

[ USS-MATLAB-FUNCTIONS ]
All matlab functions used by the example,
the main function of interest is "fit_raylsherl.m".

[ apply_uss_to_wav.m ]
The main code.


----------
More details on "apply_uss_to_wav.m"

The idea is to show the original magnitude spectrograms of a given WAV file,
then to show how the ( Rayleigh + Shifted Erlang ) distribution is fitted,
and finally the filtered magnitude spectrogram.

You can try it on various examples, stored in the DATA directory.
The following MATLAB lines can be directly pasted onto a MATLAB interactive command-line.


  Meeting room examples:

apply_uss_to_wav( 'DATA/av163-meetingroom-distant.wav', [5 15] );   % Limit to the first 30 seconds
apply_uss_to_wav( 'DATA/av163-meetingroom-lapel.wav', [5 15] );     % Limit to the first 30 seconds

  
  Numbers95 + Noisex examples:

  ( clean: show that the information is not lost )

apply_uss_to_wav( 'DATA/clean-NU-1028.streetaddr.wav' );

  ( stationary Lynx noise at 12, 6 and 0 dB )

apply_uss_to_wav( 'DATA/lynx12-NU-1028.streetaddr.wav' );
apply_uss_to_wav( 'DATA/lynx06-NU-1028.streetaddr.wav' );
apply_uss_to_wav( 'DATA/lynx00-NU-1028.streetaddr.wav' );

  ( non-stationary Factory noise at 12, 6 and 0 dB )

apply_uss_to_wav( 'DATA/fact12-NU-1028.streetaddr.wav' );
apply_uss_to_wav( 'DATA/fact06-NU-1028.streetaddr.wav' );
apply_uss_to_wav( 'DATA/fact00-NU-1028.streetaddr.wav' );


Note: a strange looking fit in the "clean-NU-1028.streetaddr.wav" case
is explained by some artificial data on silences in the waveform, 
most likely generated by the telephone speech/silence detector.
This effect does not appear on all other files,
and did not impact on the recognition performance in our experiments