Train data question (MLS)

[gen35]

Hi,
To ensure fairness of internal evaluation I can't use "Multilingual LibriSpeech (MLS)". I couldn't use Whisper if MLS is in the training dataset.

Is it possible to confirm that MLS wasn't used for training?

[glangford]

According to the paper, when comparing whisper to prior work, "Whisper performs well on Multilingual LibriSpeech, outperforming XLS-R" ... so it would not have been trained on MLS.

See 3.4 -
https://arxiv.org/pdf/2212.04356

[glangford]

...after the initial release, the large-v2 model was released. In the announcement of large-v2, there isn't a mention of new training data being added:

• Announcing the large-v2 model #661

For large-v3, the number of hours of training data increased but MLS is not mentioned

• `large-v3` release #1762