Datasets:
Data splits
are there any official data splits for train/val/test?
hi, no, there aren't any. Usually MLAAD is used as a train or test set in its entirety - which makes sense if you want to test generalization.
If you must have splits, I recommend training on english data and testing on non-english data (or vice versa).
Hi, wouldn't it be better to design speaker-disjoint splits (instead of language-disjoint) in order to avoid possible data leakage from the speakers?
If all speakers are exclusively real or fake, a model could learn to just memorise which speaker is what (leading to poor generalisation)
great point!
The speaker ids are (if possible) taken from M-AILABS, so there should be a number of speakers for which bona-fide and spoof instances exist.
Regardless, a train/test split w.r.t. languages/speaker/attacks does seem like a good idea.
Perfect, thank you very much for the quick replies! Have a nice day!
you're welcome!
let me know if you build something cool :-)