this post was submitted on 13 Jan 2024
579 points (95.7% liked)

People Twitter

5268 readers
1103 users here now

People tweeting stuff. We allow tweets from anyone.

RULES:

  1. Mark NSFW content.
  2. No doxxing people.
  3. Must be a tweet or similar
  4. No bullying or international politcs
  5. Be excellent to each other.

founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
[–] [email protected] 33 points 10 months ago* (last edited 10 months ago) (3 children)

The only way to train an AI voice model is to have lots of samples. As scummy as they are, neither Microsoft nor Apple is selling your voice recordings with enough info to link them to you specifically. This person probably just forgot about an old social post where they talk for enough time for a model to be trained. Still super scary stuff.

[–] [email protected] 34 points 10 months ago* (last edited 10 months ago)

Not true anymore. You can create a reasonable voice clone with like 30 seconds of audio now (11labs for example doesn't do any kind of authentication). The results are good enough for this kind of thing, especially in a lower bandwidth situation like a phone call.

[–] [email protected] 13 points 10 months ago* (last edited 10 months ago)

This person probably just forgot about an old social post…

Or recordings made during customer service calls, maybe a disgruntled employee decides to repurpose the data.

[–] [email protected] 6 points 10 months ago* (last edited 10 months ago)

True for creating voices at all, but that work has already been done.

Now we're just taking these large AI's trained to mimic voices and giving them a 30 second audio clip to tell them what to mimic. It can be done quickly and give convincing results especially when hidden by the phonecall quality.