Does whisper-large-v3 improve upon faster-whisper-large-v3 reproductibility wise?
#14
by RodrigoM-CGI - opened
Hi, I have a transcription project where I was using faster-whisper-large-v3 for short term demo testing, with planning to scale up to whisper-large-v3 once complete. However in my testing of longer audio files (~20 mins) I started noticing inconsistencies with the timestamps and phrases in between runs of the script. I first thought it was just cuda vs cpu processing discrepancies, but running several times on the same gpu gave me different outputs of text and timestamps, which cascade the errors downstream of the script. Would these be fixed by moving up to whisper-large-v3?