Menu

per speaker WER

Help
Horia Cucu
2011-03-28
2012-09-22
  • Horia Cucu

    Horia Cucu - 2011-03-28

    Hi,

    I have a test database composed of files that are named XX_YY_ZZZZ.mfc, where
    XX is the speaker id, YY is the recording id (phrases type 1, phrases type 2,
    etc) and ZZZZ is the file id.
    As explained in the tutorial I've created a list of all the test files
    (files_test.fileids) and a list of all the transcriptions
    (files_test.transcription) and I'm using these lists to do the decoding. I'm
    getting a Sentence error and a WER.
    Is there any way of getting the WER per speaker or the WER per recording type
    instead of the overall WER? At this point I'm creating separate lists of
    fileids and transcription for every test that I need to do, but that's really
    time consuming when you have more than 10 speakers in the testing database...

    I'm thinking to go into the internals of ./scripts_pl/decode/slave.pl and
    modify the part that computes the WER (the decoding can be done using the full
    lists, but the word aligning and WER calculation has to be split into several
    parts). Do I need to do that or is there a simpler way of computing the per
    speaker WER?

    Horia

     
  • Nickolay V. Shmyrev

    You can install NIST sclite tool and configure it in sphinx_decode.cfg instead
    of builtin word_align.pl. sclite gives much more comprehensive statistics.

     
  • Horia Cucu

    Horia Cucu - 2011-03-29

    Thanks!

    It was really simple to integrate sclite within sphinx system :). Now I only
    have to learn how to use sclite to compute WER on groups of files. And I also
    have to modify etc/decode/slave.pl to print out the various statistics.

     

Log in to post a comment.