Spaces:

competitions

/

CryCeleb2023

like 34

Paused

Please check out my solution for 1st place here: https://dienhoa.github.io/dhblog/posts/cryceleb.html

It's about using a bigger model than the default ECAPA model with n_mels = 150, lin_neurons = 250 (I decided not to use a pre-trained model as the cry sound is very different, so maybe we don't have a lot of benefits for transfer learning here )

Using Label Smoothing in loss function with value 0.05.

I have to admit that running with different seeds gives me very different results, so it means that I still don't have a very robust approach.