Bug in wave normalization. about crepe HOT 7 CLOSED

@adarob @justinsalamon Thanks for the heads up! Just pushed v0.0.11 :)

Btw you can use:

git+https://github.com/marl/crepe.git

in requirements.txt to pull the latest commit from Github, or:

git+https://github.com/marl/crepe.git@7f0bf0593847f38df47d2760934d585ad2f618c0

for a specific commit.

from crepe.

Thank you for the detailed report! This is a great catch, and it's striking that the code works regardless.

I'll push the fix shortly.

from crepe.

Just a quick question, was the model trained on audio normalized the old / incorrect way?

I am using a PyTorch port of this repo with your weights, and am wondering whether I should change or update the normalization code? Or if the model requires re-training with the fixed normalization?

from crepe.

The pre-processing code I used for training is different from this, which did not have the normalization bug in this issue. So it shouldn't require retraining and you can use the same pre-trained model, unless you want to fine-tune the model with some other dataset.

from crepe.

@jongwook will you be releasing an updated pip package with this fix?

from crepe.

Yes :) @jongwook let me know if you're slammed right now and I can make a release. Cheers!

from crepe.

Just a quick question, was the model trained on audio normalized the old / incorrect way?

I am using a PyTorch port of this repo with your weights, and am wondering whether I should change or update the normalization code? Or if the model requires re-training with the fixed normalization?

@angusturner
Hi, I‘m looking for the pytorch version of CREPE, but found nothing. Does the "PyTorch port" you mentioned is public available? or would you mind about releasing it? Thank you very much in advance.

from crepe.