What was the exact format during stage 1 and stage 2

[choyakawa]

What special tokens were used here? Could you please share one example?

[darkacorn]

that looks like the regular tokenizer .. as it produces audio tokens and text tokens at the same time - its always 12.5 audio tokens for 1 sec of audio from what ive seen so far

[choyakawa]

I mean the exact usage of those unlisted special tokens, <|begin_of_transcription|>, <|end_of_transcription|>, etc.

[darkacorn]

transcript is audio in from the whisper backbone .. the tokenizer is pretty easy