Skip to content

Silence yields random text for certain models. #19

@thewatchacker

Description

@thewatchacker

Version 0.4.1 on Windows 11 inserting into notepad.

On some models, if I activate continuous listening and then a second later I deactivate it, then random strings get inserted in multiple languages. Note as well that... If I wait longer before deactivating, I still get these random strings. Although, interestingly, I'll see one in the preview bubble, but then if I wait to press to deactivate listening, then a different string in a different language might get inserted.

So, if I do start speaking and I wait for a while, then I don't get the random word insertions. In fact, it looks like as long as I do say something, then the random hallucinated words completely disappear.

examples where In a quiet, enclosed room, I activated and deactivated. Continuous listening several times. On the GT40 example it is easier to see where the activations and deactivations were because the language changes each time. Note that in all cases I have translation turned off.

Whisper Large V3 Turbo.
So.soSoSoI'm not too satisfied.you

GPT-4o transcribe
Vorsicht!。فإنه사실은.そうだね。Ok.

After doing this enough times with the GPT model, I did have type whisper crash. I have not been able to reproduce that.

I don't see this with every model. Whisper Large V3 and GPT 4o Mini don't show this behavior for me. Also, the downloaded models such as Parakeet don't seem to do this either.

Note this could also be considered a Ouija board type feature rather than a bug since as long as you do say something then everything works fine. I will defer to the project owners.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions