ChatGPT cloned my voice and replied to herself

[@damianlewandowski2315]

I encountered it today during my English lesson, and I have to admit, it creeped me out at first. 😅

[@jsalsman]

Hahah! Anyway, advanced voice mode does not use text to speech. The output tokens are translated to both audio and text.

[@bass-tones]

By the way, it’s hallucinating when it’s explaining to you how it works. This new mode does _not_ work using text to speech or vice versa. The transcripts you see in the chat are from a completely different model, I’m assuming Whisper, transcribing the conversation after the fact into text for your own benefit. There is no text step in the actual conversation. Cool that you captured this bug either way though. I’m starting to understand better now what “safety” actually means with these things. Wild it can clone your voice despite that not being its goal and with a tiny sample of your voice to even work with.

[@JohnsonNong]

holy shit

[@jonathan.ijzerman]

It is not lying, the text generation is done by GPT4o, which doesn't have full/accurate knowledge on all auxiliary systems like advanced voice

[@Napalmbethsux]

So you're saying that the chat GPT explained the voice to text function in a lot of gobblegook double speak, closing the explanation qith, "I hope that clears things up" and then chatgpt answered its own question by saying, "yeah it does"?

[@robbrown2]

@@Napalmbethsux Yes, that's exactly what it did. Really weird that it was trying to convince me of something that I was not convinced of and it inserted the reaction it was hoping for.