Menu

Hindi (Devanagri) accuracy and no recognition on live speech

Help
2017-02-23
2017-02-23
  • Prince Dhanwan

    Prince Dhanwan - 2017-02-23

    Sir,
    I have trained a .cont model for a single speaker and I have finally reduced the word error rate to 24.9%,
    but when model is used for recognition it is giving very less accuracy,
    when using this model for recognition it is picking up the noise from the surrounding and the accuracy of model becomes very less, and in the FAQ it is mentioned that if any kind of noice reduction is applied on noise than the speech spectrum will be disturbed and feature extraction will be disturbed,
    So sir how to deal with that noise?

    http://cmusphinx.sourceforge.net/wiki/faq

    And in FAQ it is also mentioned that there is a feature remove_noise in sphinx train, but I cannot find it in sphinx train, and I am using the latest version of availavble on Github.
    And is there any possible way to increase the accuracy on live mic?

     
  • Prince Dhanwan

    Prince Dhanwan - 2017-02-23

    Here is the log with word error rate of 24.9%

     
    • Nickolay V. Shmyrev

      So sir how to deal with that noise?

      Get noise cancelling microphone

      Here is the log with word error rate of 24.9%

      Your data is not sufficient much less than recommended by tutorial.

       
  • Prince Dhanwan

    Prince Dhanwan - 2017-03-04

    can i apply a noise filter for live speech instead of a noise cancelling microphone

     
    • Nickolay V. Shmyrev

      It will not help much. Noise cancellation corrupts speech and reduces speech recognizer accuracy.

       

Log in to post a comment.