Menu

MLLR adaptation

Help
hiyassat
2011-05-07
2012-09-22
  • hiyassat

    hiyassat - 2011-05-07

    Hi
    I am trying to build an acoustic model for large vocabulary corpus (for Arabic
    language )
    Corpus description :
    3000 hr of audio file (wav file 16000)
    73,000 words (22,000 unique words)
    Number of speakers 150 speakers
    Each speaker speaks the 22,000 word (or so)
    Scenario 1) The system will be trained for some reader,. Then, the MLLR
    adaptation algorithm is applied to the remaining readers
    Scenario 2) to train the system with all readers at once
    Which one is better

     
  • Nickolay V. Shmyrev

    Best scenario is 3)

    You try to compare both 1) and 2) on smaller database (100 hours) and report
    the results here. Then you apply the best of 1) and 2) to your large database.

     
  • hiyassat

    hiyassat - 2011-05-08

    thank you Nickolay
    i will do, and i will let you know

     

Log in to post a comment.