search

marl / crepe

CREPE: A Convolutional REpresentation for Pitch Estimation -- pre-trained model (ICASSP 2018)
https://marl.github.io/crepe/
MIT License
1.12k stars 160 forks source link

Remove Dirty Training Data of Piccolo #71

Open Daniel-Chin opened 3 years ago

Daniel-Chin commented 3 years ago

Your paper mentioned some dirty training data of Piccolo that led to low test-time accuracy for Piccolo sounds. Do you think removing the problematic training data and training the model again will be a good idea?

The current model really learned to do two things: A) estimate pitch, B) if the instrument is classified as Piccolo, do a weird thing.

If we fix the dataset, the model won't have to learn a wrong thing anymore, which probably benefits its performance on other instruments too. Maybe then you will be able to decrease the model size without sacrificing accuracy?

Daniel-Chin commented 3 years ago

Since it is still SOTA (as of 2020), improving the model can be beneficial to many ;-)