Training data on model card different from paper?
#2
by
AbrahamSanders
- opened
Hello, and thank you for sharing the EnCodec model!
I noticed the model card indicates that this model (the 48kHz stereo EnCodec) was trained on speech, music, and general audio data. However, the paper (https://arxiv.org/pdf/2210.13438.pdf) indicates that the 48kHz stereo model was only trained on music data:
4.1: Dataset
We train EnCodec on 24 kHz monophonic across diverse domains, namely: speech, noisy speech, music and
general audio while we train the fullband stereo EnCodec on only 48 kHz music.
It appears the model cards are almost identical between the 24kHz mono and 48kHz stereo models - perhaps this difference can be documented here for the 48kHz checkpoint, to reduce potential confusion.
All the best!