-
Notifications
You must be signed in to change notification settings - Fork 147
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Junk output when not spoke or for some junk sound. #14
Comments
@BakingBrains That's a known issue with a lot of AI models, called hallucination. Not much can be done about it unfortunately at this stage afaik |
Yep @SunnyOd , I knew, I thought there might be some method so that we can suppress the junk output. But thanks though👍 |
Perhaps some analysis could be done on the wav audio to see if its silence or not |
Hi. Has anyone managed to suppress the junk output? I keep getting "Thank you" and a couple of other phrases pop up during inactivity/silence. I've played with energy levels and that has helped some but I think there might be more mileage with the fix below - possibly I've found some threads where people have managed to fix it, this one looks to give the solution as including Thanks! |
hi ! it's possible to get more information about the context of your bug, like the program or somethings else ?Thanks Boulanger / Melvin Gueneauenvoyé : 11 août 2023 à 22:11de : SunnyOd ***@***.***>à : mallorbc/whisper_mic ***@***.***>cc : Subscribed ***@***.***>objet : Re: [mallorbc/whisper_mic] Junk output when not spoke or for some junk sound. (Issue #14) Hi. Has anyone managed to suppress the junk output? I keep getting "Thank you" and a couple of other phrases pop up during inactivity/silence. I've played with energy levels and that has helped some but I think there might be more mileage with the fix below - possiblyI've found some threads where people have managed to fix it, this one looks to give the solution as including --suppress_tokens in the command when running whisper. Not sure how to add this flag to whisper_mic since the move to pip installation of whisper_mic. Is it worth looking into?Thanks! S—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
|
So this is due to hallucinations. There doesn't seem to be a way to fully fix it due to the way that the model was trained. It hallucinates when there is little to no audio so it makes something up. I added a new flag that helps with this. I'm going to leave this issue open. If anyone finds a real solution, please ping me here. However, I think with the current flags that exist with the tool that you can find a configuration that limits hallucinations. |
Thanks for the Cool Repo. But I found when there is no audio the model gives junk transcription.
Any suggestions to improve this here?
The text was updated successfully, but these errors were encountered: