Menu

Regarding system performance

Help
SKR
2017-06-02
2017-06-02
  • SKR

    SKR - 2017-06-02

    Hi all,,

    Greetings..

    I selected 10 speakers, told them to speak 30 words. Then I select wave files of these 9 speakers to train the acoustic model and the remaining one set wavefiles for testing....

    I got 100% WER and SER during decode..... !!

    I surprise,, the engine made to study the same thing from all speakers but not able to recognize the same thing by a different speaker .......

    How it possible.....? Is sphinx purely dependent on speakers??

     
    • Nickolay V. Shmyrev

      You can share your training folder to get help on this issue.

       
  • SKR

    SKR - 2017-06-02

    Of course.. Here is a sharable link to my data....

    https://drive.google.com/file/d/0B_74UylilDfCYmdyZmhhYlJSaEE/view?usp=sharing

    Thank u for response..

     
    • Nickolay V. Shmyrev

      Your mistake is that arpa model is not properly prepared, it is build from phones instead of words.

      If you build arpa lm from words, it will be much more accurate, it will guess most of the words actually.

       
  • SKR

    SKR - 2017-06-02

    Okay,,, That was a mistake.... Thank uu...

    :)

     

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.