Menu

Low accuracy when testing after training acoustic model

Help
2019-05-15
2019-05-15
  • Le Gia Anh Quy

    Le Gia Anh Quy - 2019-05-15

    Hi, I've followed tutorial https://cmusphinx.github.io/wiki/tutorialam/ to train a new acoustic model for my Vietnamese language. Simply, I've recorded digits from 0 to 9 to train and also used these trained data to test. Therefore, the accuracy should be 100% as expected. However, it's only about 20% (only recognize 2 and 3). I've re-recorded many times but it's not changed. I can make sure that I do exactly all requirement such as 16k sample rate, 16 bit depth,...

    Could anyone give me some explanation and the way to improve my model? Thank y'all

    Here is my project: https://github.com/legiaquy/test

    Note: In 'sphinx_train.cfg', I changed 2 lines for a small amount of data:
    $CFG_CD_TRAIN = 'no'; [165]
    $DEC_CFG_MODEL_NAME "$CFG_EXPTNAME.ci_cont"; [242]

     

    Last edit: Le Gia Anh Quy 2019-05-15
  • Le Gia Anh Quy

    Le Gia Anh Quy - 2019-05-15

    Well, I've found the solution myself. For anyone have the same issue, I rebuild my language model using CMUCLTK instead of using web-service LMTool

     

    Last edit: Le Gia Anh Quy 2019-05-15

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.