Improving SincNet results on TIMIT by adding reverberation

[seandickert]

Not an issue, but just wanted to post that you can further decrease the sentence classification error by reverberating each training call and including them in the training dataset (effectively doubling the training size). The error drops to 0%. I will try with Libri as well

[mravanelli]

Interesting...

On Feb 26, 2019 20:44, "seandickert" notifications@github.com wrote:

Not an issue, but just wanted to post that you can further decrease the sentence classification error by reverberating each training call and including them in the training dataset (effectively doubling the training size). The error drops to 0%. I will try with Libri as well

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mravanelli/SincNet/issues/18, or mute the thread https://github.com/notifications/unsubscribe-auth/AQGs1lyoCdTRFLsKDpHzXVtkeoTN-h0eks5vReMMgaJpZM4bTgw5 .

[mingewang]

Oh , cool, just wondering how to doreverberation? Any Python lib? Thanks

[seandickert]

Mr Ravanelli (author of the paper and this repo) has another repo which will convolve a signal w/ an impulse response. There are also a few IR files in that repo. I used #1. Repo: https://github.com/mravanelli/pySpeechRev

[mingewang]

@seandickert , Thanks! BTW, do you see the similar result in libri data? have you tried real audio data?

[hdubey]

@seandickert When do you perform reverberation, after reading the file chunk or do it during data processing stage? Can you share you code snippet. I guess TIMIT results were already good, 0% is FER or Sentence error rate ? Did you tried adding noise as well. Thanks!