Modern models need about 5sec of audio to replicate a voice. The days when you needed a large amount of audio for replication are long gone. Same for faces by the way, the original Deepfake needed hundred of images and hours of training, now you can do it with as little as a single good image instantly. Software to automatically clone the voice, translate the audio into another language and adjust the lip motion exists as well, again without any lengthy training or material, just needs the clip you want to change.
Where the whole thing gets interesting is in remixing. If you stolen Stephen Fry, sure that’s bad and there might be laws against it. What if you remix Stephen Fry and Patrick Steward into a brand new AI-persona? What if you remix a Stephen Fry sound-alike out of other peoples voices without ever touching his voice?
Do you have a reference for the 5secs of audio? And what would the quality of the ai be? I mean it would have to infer a lot of training just given 5 seconds of audio.
Modern models need about 5sec of audio to replicate a voice. The days when you needed a large amount of audio for replication are long gone. Same for faces by the way, the original Deepfake needed hundred of images and hours of training, now you can do it with as little as a single good image instantly. Software to automatically clone the voice, translate the audio into another language and adjust the lip motion exists as well, again without any lengthy training or material, just needs the clip you want to change.
Where the whole thing gets interesting is in remixing. If you stolen Stephen Fry, sure that’s bad and there might be laws against it. What if you remix Stephen Fry and Patrick Steward into a brand new AI-persona? What if you remix a Stephen Fry sound-alike out of other peoples voices without ever touching his voice?
This whole issue gets very blurry very fast.
Do you have a reference for the 5secs of audio? And what would the quality of the ai be? I mean it would have to infer a lot of training just given 5 seconds of audio.
Here is an alternative Piped link(s):
Translation with HeyGen
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source, check me out at GitHub.