Training data information

[marisbasha]

Hello @jongwook, I wonder if you could provide information if Whisper large-v2 was trained with the Ami Speech Corpus and TIMIT. I am writing a paper on using whisper embeddings as a learned similarity, and I need to know this information since if it's pretrained on those datasets it would make it hard to make assumptions.

[jongwook]

Hi, we didn't specifically include those two datasets in training, but some samples from them may have been mixed into the data given the scale of the dataset. I'd still consider them "out-of-distribution".

[marisbasha]

Thanks a lot!