Anyone know how many hours of training data it takes to build up a convincing model of someone’s voice? It was 10’s of hours when I did a bit of research a year ago… the article says social media is the likely source of training data for these scams, but that seems unlikely at this point.
The technology has clearly come a long way in a short time, really fascinating.
I remember the first examples I read about being trained with celebrity read audiobooks because they needed so much audio data. I want to say Tom Hanks or Anthony Hopkins but I could have that confused with something else.
A current state of the art ai model from Microsoft can achieve acceptable quality with about 3 seconds of audio. Commercially available stuff like eleven labs about 30 minutes. But quality will obviously vary heavily but then again they’re using a low quality phone call so maybe not that important
That’s downright scary :-) I think it took longer in the last Mission Impossible.
30 minutes is still pretty minimal for the kind of targeted attack it sounds like this is used for. I suppose we all need to work with our families on code words or something.
I went in thinking the article was a bit alarmist, but that’s clearly not the case. Thank for the insight.
With that little, they may be able to recreate the timbre of someone’s voice, but speech carries a multitude of other identifiers and idiosyncrasies that they’re unlikely to get with that little audio, like personal vocabulary (we don’t choose the same words and phrasings for things), specific pronunciations (e.g. “library” vs “libary”), voice inflections, etc. Obviously, the more training data you have, the better the output.
Anyone know how many hours of training data it takes to build up a convincing model of someone’s voice? It was 10’s of hours when I did a bit of research a year ago… the article says social media is the likely source of training data for these scams, but that seems unlikely at this point.
deleted by creator
The technology has clearly come a long way in a short time, really fascinating.
I remember the first examples I read about being trained with celebrity read audiobooks because they needed so much audio data. I want to say Tom Hanks or Anthony Hopkins but I could have that confused with something else.
A current state of the art ai model from Microsoft can achieve acceptable quality with about 3 seconds of audio. Commercially available stuff like eleven labs about 30 minutes. But quality will obviously vary heavily but then again they’re using a low quality phone call so maybe not that important
That’s downright scary :-) I think it took longer in the last Mission Impossible.
30 minutes is still pretty minimal for the kind of targeted attack it sounds like this is used for. I suppose we all need to work with our families on code words or something.
I went in thinking the article was a bit alarmist, but that’s clearly not the case. Thank for the insight.
With that little, they may be able to recreate the timbre of someone’s voice, but speech carries a multitude of other identifiers and idiosyncrasies that they’re unlikely to get with that little audio, like personal vocabulary (we don’t choose the same words and phrasings for things), specific pronunciations (e.g. “library” vs “libary”), voice inflections, etc. Obviously, the more training data you have, the better the output.
TorToiSe can work off of just three ten second clips when you’re using a pre-trained model. No telling if that’ll sound any good.
I’ll have to check that out, thanks for the link.