Random words recognized while not speaking or speaking non-English

[rudymohammadbali]

I am facing an issue with the transcription of audio using the Whisper module. It randomly recognizes words like "Thanks you" and "Thanks for watching" even when I'm not speaking or speaking in a language other than English.

Has anyone else encountered this problem? If so, I would greatly appreciate any guidance or suggestions on how to resolve this issue.

I have tried adjusting the settings and parameters of the Whisper module, but I haven't been able to find a solution. It seems to be an issue with the speech recognition algorithm incorrectly interpreting non-speech sounds or capturing random words.

Is there a way to fine-tune the module to improve its accuracy and avoid recognizing irrelevant words? Or perhaps there are specific settings or techniques I should consider to mitigate this issue?

Any help or insight into this problem would be highly appreciated. Thank you in advance for your assistance!

Here is a example of what it recognized while not speaking or speaking in other language:

Nice, DJ. Thank you. AHHH! AHHH! What? Thank you. Thank you. Thank you. Thank you. a lot of the function if you're fancy to any The better my trip will get. Thank you. Thank you. I know, probably. Put it in. Put it in. Thank you. Thanks for watching! I'll see you later. I'll see you next time. Thanks for watching! Thanks for watching! thank you Thank you. I'll see you next time. Thank you. Thank you. Thanks for watching! Thank you. Now Quyyy Comment and Subscribe A little bit of history, right? so good it's funny English Story. Write me. I get you. Discussion. For the health of wise people. But look, here's a problem. Here's a problem. Take me right down the road. But at night. While not speaking. Description. Description of the Lord. Ava. Almighty. Let's just spend the eveningbreak Mincey cantat Aw All right What is it? Universal. I think it's 10 now. I'm not allowed. I'm shit. One, one. I'm not allowed. I'm not allowed. It is a Need for Humanity. I'm not sure how I'll be able to do it. I don't know if I'll be able to do it. I don't know if I'll be able to do it. I don't know. I don't know. I'm not sure if I can do it. I'm not sure if I can do it. Welcome! Hey! I can do it with Chef D, too. You can do it with Chef Sun. Ah. Thank you. I'm not sure what to say. I'm not sure. Do you have any questions? I don't have any. I don't have any. I don't have any. I don't have any. For real... Four hands since from there I'll see you next time. But it's not like there's thattheme I want two big keys! So it all there Thanks for watchinglaughs We brought Ronald, Is this the first time? Yes, the first time. I think it's the first time. I'm not sure if it's the first time. so se files are fully deployed? Yeah, they are pre baking sheets and they still weren't as That's what it feels like. Thank you. I'll see you next time. Thank you. Thank you. I'll see you next time. I'll see you next time. Thank you. Please see the complete disclaimer at PissedConsumer.com I'll see you next time. I'll see you next time. I'll see you next time. I'll see you guys next time. I'll see you guys next time. I'll take a look. Do you want a bit? Is that cool? Yeah. Thank you. Okay, excuse me. You're welcome. Okay. I'll go and look for you. I'll go and look for you. Thank you. We'll just let him go. Thank you. Thank you. Richard Peaceful… Thank you. Maybe… Thank you. I don't. Good stuff. I'll see you next time. Thanks for watching! Thank you for watching!

[phineas-pta]

it's the hallucinations problem, happened when there's silence / non speech segments

see #679

[rudymohammadbali]

I did not find a solution. Please tell me the answer if you know it and thanks.

[glangford]

see also

• Dataset bias ("❤️ Translated by Amara.org Community") #928

[rudymohammadbali]

If no fix from OpenAI so I can do nothing but else wait for a solution.

[ryanheise]

Waiting for OpenAI is not the only option, as many people have shared solutions to the hallucination problem on this discussion board. Now that you know what the problem is called (thanks to the people who answered you), you can try searching for that keyword here.

[rudymohammadbali]

I'm making real-time speech recognition. so the recorded audio file is about 2-3 seconds. and I don't know how to process this audio file to remove the silent parts.

[misutoneko]

You might want to take a look here:
#1488

So basically it seems like a bug somewhere in how non_speech_tokens is handled.
I haven't received much feedback on that btw, so all I can say about it is that it works for me.
My samples are also very short, so maybe it helps you too.

[rudymohammadbali]

thanks for answer.

[MiguelQueiroz]

yeah, I use a voice assistant that uses openAI whisper for speech detection/commands, and when there is silence I get a response back talking about Review No.106161 on PissedConsumer.com or other type of review but it mostly goes to PissedConsumer.com most of the time dont know why! This is because the voice transcription was detected as such! 😩