Menu

FSG gives more recognition errors than n-gram

Help
2011-05-08
2012-09-22
  • Pranav Jawale

    Pranav Jawale - 2011-05-08

    Hi,

    I am using sphinx3.8

    I was comparing performance of FSG and unigram LM on a 5 word vocab. All the
    words are equiprobable, all my test files contain just one of the words. I
    found that with FSG I can force the o/p to be a single word, but on the other
    hand in terms of # words wrongly recognized (I had 15 test wav files for each
    word), FSG performed poorly. i.e. even though with unigram occasionally 2
    words were recognized, # words recognized correctly with unigram was much
    greater than that with FSG.

    Could someone please explain / point to some reference as to why might FSG be
    poorly working as compared to n-gram? Is FSG / n-gram search explained
    anywhere?

    Thanks a lot.

     
  • Nickolay V. Shmyrev

    Hello

    It depends on many things. I suggest you to share your test so we can take a
    look

     
  • Pranav Jawale

    Pranav Jawale - 2011-05-08

    Thanks for offering to help!

    Please find my test here http://home.iitb.ac.in/~pranavj/FSGvsN-
    GRAM_test.zip

    Let me know if there is any problem with the zip file.

    Please read README.txt in it. It contains info about all the files present in
    the folder. I am running windows compiled version of sphinx3 btw.

    Results of my test were -
    Unigram - # files Correcttly recognized (out of 142) = 119
    FSG - # files Correcttly recognized (out of 142) = 92

     
  • Pranav Jawale

    Pranav Jawale - 2011-05-08

    In case you get any error like # phones in FSG search has to be <= 64, please
    change the appropriate param in source file (it is explained in the error
    msg). I wonder why this # phones limit has been imposed.

     

Log in to post a comment.