Menu

How is adaption accuracy measured?

Help
Hitarth
2012-08-29
2012-09-22
  • Hitarth

    Hitarth - 2012-08-29

    Hi,

    On the "Adapting the Default Acoustic Model" page at http://cmusphinx.sourcef
    orge.net/wiki/tutorialadapt,
    close to the bottom of the page I found the following stats:

    TOTAL Words: 773 Correct: 669 Errors: 121
    TOTAL Percent correct = 86.55% Error = 15.65% Accuracy = 84.35%
    TOTAL Insertions: 17 Deletions: 11 Substitutions: 93

    Now the first stat is pretty obvious as it is just the total number of words
    to be recognized. However when we get to "Correct" and "Errors", does being
    correct mean that the word not only gets recognized correctly BUT also in the
    right order? and if it doesn't get recognized in the correct order in which it
    was input, does that get classified as an error?

    Moving on to the second line, it was pretty easy to figure out how % for
    Correct and Error was derived. However, I couldn't figure out how accuracy %
    was derived? and why aren't Correct and Accuracy % the same?

    Finally, on the last line, insertions, substitutions and deletions are
    mentioned. What do those stats mean though? I can guess that maybe
    substitutions shows how many words were in the wrong order? But I have no clue
    about insertions and deletions.

    I will really appreciate any help understanding all this.

    Cheers

     
  • Hitarth

    Hitarth - 2012-08-29

    Btw this output was derived using script word-align.pl from Sphinxtrain (as
    specified at the url in the previous post)

     
  • bic-user

    bic-user - 2012-08-29

    accuracy = (total words - insertions - deletions - substitutions) / total
    words.
    You can read this http://en.wikipedia.org/wiki/Levenshtein_distance a bit to get where from I, D, S
    comes

     

Log in to post a comment.