Transcribing translates audio into English. #186
-
I have an audio file with a speech in Russian. Somehow the whisper (large model) detects the English language and using the basic parameters the model automatically translates all speech from Russian into English when transcribing. If I specify the language as Russian than the model skips the first 15-25 seconds of the audio. The attempts of manually specifying the task as 'transcribe' didn't help either. Any ideas how to fix it would be appreciated. Thanks! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
Looks like our language detector was not accurate enough, and specifying Russian is the intended approach in this case. Tweaking the |
Beta Was this translation helpful? Give feedback.
Looks like our language detector was not accurate enough, and specifying Russian is the intended approach in this case. Tweaking the
max_initial_timestamp
option may help in this case, but otherwise I'm not sure how exactly we can fix it. If you can upload/email the audio file, I can try looking into it.