[ file MULTISEG/README ]

Contact: lathoud@idiap.ch

===========
DESCRIPTION
===========

Tool to segment multichannel recordings (e.g. lapels or headsets):

1) You have recorded several WAV files (one for each person),
   synchronized in time.

2) You can use the MATLAB function "segment_multichannel.m"
   to produce a single ".txt" file containing a list of speech segments
   for each channel.

3) You can use the scripts in the TXT2XML directory
   to convert the ".txt" file into XML format (see TXT2XML/README).

This is an implementation of the baseline multichannel segmentation
presented in a paper at the 2004 NIST Meeting Recognition workshop (RT-04),
titled "Unsupervised Location-Based Segmentation of Multi-Party Speech".

MATLAB files written by Guillaume Lathoud (lathoud@idiap.ch).
TXT2XML tools written by Maël Guillemot (mmmAdmin@idiap.ch).


==========================
CONTENTS OF THIS DIRECTORY
==========================

DATA/                      -> an example (WAV files of one short meeting).

README                     -> this file

TXT2XML/                   -> additional script to convert the ".txt" output
                              of "segment_multichannel.m" into XML format 
                              (see TXT2XML/README)

check_param.m              -> MATLAB script used by "segment_multichannel.m"
dilation.m                                (idem)
energy.m                                  (idem)
erosion.m                                 (idem)
fill_default.m                            (idem)

segment_multichannel.m     -> the main segmentation script


==========
AN EXAMPLE
==========

% Some useful MATLAB lines & information.

% Advice: do not use MATLAB 6.5
% Use MATLAB 6.1 instead

% Information
% This prints most of the information you need (also copied at the end of this README file).
% Whatever is missing can be found by directly looking at the code:
%   dbtype segment_multichannel
% or drop me an e-mail.

help segment_multichannel;

% Example of use

clear in_p;
in_p.wavfile_list{ 1 } = 'DATA/Lapel-1_T000010.0_T0459.201.wav';
in_p.wavfile_list{ 2 } = 'DATA/Lapel-2_T000010.0_T0459.201.wav';
in_p.wavfile_list{ 3 } = 'DATA/Lapel-3_T000010.0_T0459.201.wav';
in_p.wavfile_list{ 4 } = 'DATA/Lapel-4_T000010.0_T0459.201.wav';
in_p.out_dir = 'DATA';
in_p.output_basename = 'myMeeting';
[ multiseg, p ] = segment_multichannel( in_p );

% Once this is done, you will obtain a file
%
% 'DATA/myMeeting-multiseg.txt'
% 
% as well as several intermediary '.mat' files.
% Keep the '.mat' files if you want to play with parameter
% of "segment_multichannel", otherwise you can delete them.


================
MORE INFORMATION
================

>> help segment_multichannel 

  FUNCTION [ MULTISEG, P ] = SEGMENT_MULTICHANNEL( IN_P )
 
  Segment a multichannel audio recording (e.g. lapels or headsets).
  Produces intermediary ".mat" files as well as a human-readable
  "-multiseg.txt" file. The output is also returned in the form
  of a struct array: MULTISEG (see below).
 
  This is an implementation of the baseline multichannel segmentation
  presented in a paper at the 2004 NIST Meeting Recognition workshop (RT-04),
  titled "Unsupervised Location-Based Segmentation of Multi-Party Speech".
 
 
  IN_P is a structure containing parameters, of which two 
  are mandatory:
    IN_P.WAVFILE_LIST = cell array of NCHANNELS strings (filenames).
    IN_P.OUTPUT_BASENAME = string, the base name for all output files.
 
  There can be many other parameters (look at the code,
  e.g. with "dbtype  segment_multichannel").
 
  MULTISEG is a struct array of NCHANNELS elements.
  MULTISEG( i_channel ).SEG is a 2-row matrix. Each
  column describes a speech segment (row 1 = start time in seconds,
  row 2 = end time in seconds).
 
  P is IN_P enriched with default values, 
  in particular it gives the names of the output files.
 
  
  NOTE: for some reason this script does not work well with MATLAB 6.5.
  Use MATLAB 6.1 instead.
 
  By Guillaume Lathoud - lathoud@idiap.ch