> It's kind of broken, but it took zero effort.
What does "zero effort" mean? I'd be surprised if you told me that Tupac's voice was recreated without using any copyrighted material.
What does "zero effort" mean? I'd be surprised if you told me that Tupac's voice was recreated without using any copyrighted material.
This fascinates me and opens up a wholly new can of worms.
Technically, it's mostly Linda Johnson's speech data, which is in the public domain. Most voice models incorporate this as the base for learning how to transform phoneme to spectrogram, spectrogram to waveform. They'll then transfer learn on top of this.
Could you tell the difference between Tupac and an impersonator? What about 10% Tupac, 90% impersonator?
What if it's really 1% Tupac, 9% impersonator, 90% Linda Johnson? Or no Tupac at all? Can anyone tell the difference?
Where did the copyrighted data go? Is there any left?
Probably not, but learning to impersonate Tupac isn't "zero effort" (and I'd imagine that it required access to copyrighted material).
To be explicit: copyright is awful and I'd like to see it abolished. It's important to realize that nifty stuff like Tupac vocoders are not "zero effort" and require public access to data that intellectual propertarians want to have monopolies on.