ICSI Speech FAQ:
7.5 How can I use the alignment files produced by the SRI trainer?

Answer by: wooters - 2002-09-06

Here is a note from Andreas about how to deal with the alignment files produced by the SRI trainer: 
Subject: alignment file conversion
Date: Sun, 28 Jul 2002 13:14:11 PDT
From: Andreas Stolcke

the programs converting the trainer alignment file format (produced by
go.dump-alignments) from ascii to binary and back are
  
        alignments-ascii2bin
        alignments-bin2ascii
  
They are installed in $EVAL2000/bin/$MACHINE_TYPE. (Chuck Note: see
below for more info on $EVAL2000 and $MACHINE_TYPE.)
  
The output of alignments-bin2ascii looks something like this:
  
        -pau- 0 0 - 2 6 1.000400
        -pau- 0 0 - 2 7 1.000400
        -pau- 0 0 - 2 8 0.951517
        -pau- 0 0 - 2 9 0.030357
        treating 0 0 t 0 8 0.048921
        treating 0 0 t 0 9 0.970059
        treating 0 0 t 0 10 0.048077
        treating 0 0 t 0 11 0.001049
        treating 0 0 t 1 10 0.952374
        treating 0 0 t 1 11 0.999500
  
The meaning of the columns is:
  
        #1      word label
        #2      wordnet node index
        #3      wordnet arc index (out of given node)
        #4      phone label
        #5      state number (within phone)
        #6      frame number
        #7      posterior probability (of being in that state at that frame)

To find the total posterior probability for a phone at a given frame
you should sort the output by frame number and phone labels, and add
up all the probabilities with the same phone label.
The environment variable "$EVAL2000" should be set to:
/u/drspeech/opt/eval2000/system
and the $MACHINE_TYPE environment variable can be set using $EVAL2000/bin/machine-type, which on the sparcs, produces "sparc-elf" and on the linux boxes, produces "i686".

Previous: 7.4 What is forced alignment? - Next: 7.6 What is embedded training?
Back to ICSI Speech FAQ index

Generated by build-faq-index on Tue Mar 24 16:18:17 PDT 2009