I am trying to build a CNN model which should be able to identify the language being spoken in an audio file. I have extracted the MFCC matrix (for 13 coefficients) for each audio file and trained it. The audio files are call recordings. The accuracy is too high over a small data sample. The problem I have now is that in many audio files, there are common voices. I have a suspicion that the model is giving high importance to these voice features instead of features which are important to identify language (i.e. since 1 person speaks in only one language, the model might think the voice features of those files are important for classification - this will make it fail while testing for new voices). How do I manipulate the data such that each audio files has a different voice? (example: changing frequency is one way)

  • $\begingroup$ Try testing your hypothesis, feed new examples from any call you can produce $\endgroup$ Aug 17, 2020 at 13:44
  • $\begingroup$ Tested it, on new calls its gibing around 72% accuracy (dropped from around 91% before) $\endgroup$
    – user75252
    Aug 18, 2020 at 5:44


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.