Hacker News new | past | comments | ask | show | jobs | submit login

In the age of replicated AI voices and security it seems like this would be a bad idea. How much voice sample is needed to decently replicate someone's voice?



I wish I had a decent way to calibrate my paranoia about stuff like this.

I don't want to miss life's nice opportunities just because I was too worried about this kind of thing. But I don't want to be too lax either.


Well tuned zero shot models can use 5 seconds of audio, but the results aren't perfect. You won't capture prosody information, for example.

The human voice isn't as unique as you might think, though. You can encode a lot of information about a voice in about 100Kb.


Anyone could record you for this purpose at any time.


But first they'd have to sneak into my house and somehow trick my into talking.


Or have you visit a convenient website that streams your voice to the internet?


No one could be that dastardly.


Or, you know, find your LinkedIn or phone number and offer you an incredible work opportunity.


and do what with this exactly?

technically the author could record all audio here, and use it for model training or snoop into the conversation


Impersonate your voice.

Think of the old scam where someone texts you and says they're your granddaughter and they're stuck somewhere and need an urgent money transfer. In the old days the advice would be: call your granddaughter and have get confirm the story. But now with AI tech your granddaughter could call you up and deliver the scam in her own voice.


So you also don’t go to the shopping centre because they could have a mic recording you?


as it pertains to this discussion, No i don't think people are reliably snooping my voices in a shopping centre. But I do go to shopping centers.

I'd rather not sit down with a stranger without knowing them and in a format where I can be recorded. so Omegle would've been out.

That said if this site had less anonymity and maybe a registration I would be more comfortable with it.


See whisper-speech posted yesterday.


WhisperSpeech – An open source text-to-speech system built by inverting Whisper https://news.ycombinator.com/item?id=39036796




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: