README for alignSpondee

2006-10-26 Dan Ellis dpwe@ee.columbia.edu

This set of routines is for making phone alignments of spondees
(two-syllable compound words like "dogfight").  It works by making a
list of possible pronunciations from the original word (based on a
large speech recognition dictionary), then using HTK with
1ms-resolution TIMIT-trained phone models to make a forced alignment
on each, then choosing the best fit.

Usage

In MATLAB, cd to this directory, then

  alignSpondee(nameofspondeefile)

This should write a file alongside the spondee file called something
like spondee.lab, which will be in HTK label file format (with seconds
as the time units).  This can then be read e.g. by WaveSurfer to
display the alignments.

e.g. by using the included demonstration recording, 

  alignSpondee('list-e-vaseacase.wav');

writes a file "list-e-vasecase.lab", which should contain:

0.000000 0.250000 h#
0.250000 0.306000 v
0.306000 0.459000 ey
0.459000 0.595000 s
0.595000 0.640000 kcl
0.640000 0.721000 k
0.721000 0.933000 ey
0.933000 1.184000 s
1.184000 1.740000 h#

Each line is one phone label, with starttime, endtime, and (arpabet) 
phone label.

The current hmm models are very crude (single gaussian, no
re-estimation); if the accuracy is not sufficient, it should be
possible to train much more accurate models with a little more effort.

 * end *
