how can I find out log probabilities for each token/word from pretrained whisper-large-v3?
Any guidance is much appreciated.
1 Like
Hey buddy, any luck with your issue? I’m looking for the same thing without much luck lately.
Hi.
I was looking for the same thing, but couldn’t find a way with hugging face.
Instead I used whisper from OpenAI github page. The output format is different with that. Also, the default configs are different. So the output might be slightly different from what you get in hugging face.
You could also use the Faster-whisper. It’s more like the hugging face imementation and is much faster and give you word level probabilities and log-probabilities fir each segment.
Using these two is not more complicated than hugging face. So they were a goos approach for me.
1 Like