word-level-timestamps calculates promp incorrectly and disregards condition_on_previous_text (bug)

[jmward01]

Unless I am missing something, the calculation for prompt always includes at least the previous segment tokens even if condition_on_previous_text = False since prompt_reset_since is calculated before the tokens from the current segment are added. This behavior is only in the word-level-timestamps branch currently. The fix is to just move this calculation to the end like the main branch has it.

whisper/whisper/transcribe.py

Line 259 in 8eb29c3

if not condition_on_previous_text or result.temperature > 0.5:

[clementabary]

Yep, I found out while fine-tuning Whisper with timestamps but without ever giving it previous tokens. While debugging the transcribe function, it seems that with condition_on_previous_text=False it still gives it the transcription from the last segment, so technically it does not feel like "not conditioned"