Skype does it in real time (for live translation) with a few seconds of audio. For the reasons discussed in this thread, it continuously forgets its previous training from the call to not make the voice too similar, but just similar enough to distinguish the speakers.

Just a few seconds of speech gets you pretty far with a lot of models

Yes, this is generally true, in order to clone the timbre/pitch/tonality, but if you're going to run a voice cloning scam to fool friends and family members, you need much more to get the cadence, pauses, vocal tics, etc right.

Almost none, at most as little as a professional impersonator requires. GPT-4o's advanced voice mode would clone the user's voice by accident. A recording clip of one incident is available:

>During testing, we also observed rare instances where the model would unintentionally generate an output emulating the user’s voice.

>Example of unintentional voice generation, model outbursts “No!” then begins continuing the sentence in a similar sounding voice to the red teamer’s voice

https://assets.ctfassets.net/kftzwdyauwt9/4CG0G7y9WOfEkzBpi7...

https://openai.com/index/gpt-4o-system-card/