Arthur

Thanks for this comment! @narsil, I think it makes sense

Hey! For finetuning with timestamps, you should either use the latest tokenizer (which by default should add 1500 special tokens, not more) or use the previous one, which also supported...

EDIT: in order to actually solve this, we would need a lot of potential usage. The reason is that after fixing the positional ids with a `registered buffer` we need...

Hey! Thanks for pointing this out! I do agree with you on this one, I am also guessing that if you have `he said that 'revenge' was over` the same...

Hey, I think setting `devic_map = "auto"` should help balancing the load when using the `flan-t5-xxl` model to both CPU and GPU. This should allow you to run on both....

Test are failing because we do not check the `model.config` updating and will look good!

The same should now be applied to both the `TF` and the `Flax` version as the overwriting of the `generate` function is also supported. Will open a follow up PR...

For TF and flax #21334

Hey! Did you get similar results using the original `Whisper` model? Most probably this is just the model not being very good.

OK! There can be a few things to try, but if you can share an audio file with me it would help a lot!