Nicolas Patry
Nicolas Patry
@frankiedrake Have you tried using `return_timestamps` ? When experimenting I found that the default for openai is using timestamps (even if not shown) and that the model seems to perform...
`generate(..., return_timestamps=True)` to get the proper generation mode. That's correct right @ArthurZucker
`pipe = pipeline(..., return_timestamps=True)` should work though.
Without being able to reproduce it's really hard. Could you dive to the level of logits and figure out any potential differences ? I'm pretty sure it should come down...
> f I don't specify language manually the model seems not to break Can you share how you specify language in both `openai` and `transformers` ? The difference is likely...
> Maybe you can shed some light on why these parameters are so important? LMs are know to hallucinate by repeating tokens, or several tokens. I don't think there's a...
Do we agree the openai test you're doing is simply ``` whisper whisper_test.wav --model small ``` And get this ``` Detecting language using up to the first 30 seconds. Use...
Can you share anything reproducible ? Right now it's a back and forth and we can't reproduce anything on our end. Please share a clear (small) script, that I can...
I don't have the same output for neither `openai/whisper` nor `transformers` on your file.
How? This pipeline requires a feature extractor to see the document doesn't it?