Menu

trainning likelihood

hiyassat
2009-06-23
2012-09-22
  • hiyassat

    hiyassat - 2009-06-23

    i do some training for Arabic acoustic model using 73 training hour
    my dictionry contains 17000 word
    the transcription contain 100,000 words
    senones 3000
    Gaussians 16
    the obtained likelihood about 10.2
    when i do decoding using the same training data i got WER 20% and SER 55%
    what do you think ? is it OK
    is the training data enough for 17000 word in dictionary?

     
    • atizgui

      atizgui - 2009-08-16

      I do some trainig for Arabic but just aTotal Hours Training witch =1.15328803418804.
      and my dictionary contains just 26.

      But I have got some errors witch I dont know where they come like

      ERROR: "........\src\libs\libmodinv\gauden.c", line 1700: var (mgau= 117, feat= 0, density=0, component=35) < 0

      Have you any idea about this kind of error?? If I had this error, my model wouldnt work with sphinx4??

      Thank you for any replay or response.

       
    • Nickolay V. Shmyrev

      Mostly fine. I'd use 4000 senones, the rest should be tuned for this particular database. Things like forced alignment should or vtln por proper language weight should give few more percents.

       

Log in to post a comment.