At a certain point with generative AI we're going to run out of voices and faces the same way we run out of domain names and trademarks.

▲

b00ty4breakfast 5 hours ago | parent | next [-]

Aren't these models are trained publicly available data? this might hold for some rando you doesn't have their likeness in many places to be gobbled up by the Datamongers but these programs imitating someone who has been in the media for 20 years like David Greene is not the result of chance unless you are being excessively charitable.

Even if it is complete chance, there's no way to peer inside and confirm that because these things are completely opaque black boxes

▲

kelseyfrog 6 hours ago | parent | prev [-]

Can we not sample indefinitely from the latent space of vocal and delivery characteristics?

▲

parpfish 5 hours ago | parent [-]

the "latent space containing all voices" may give you the ability to parametrize voices and make an infinite number of unique voices. BUT... people have a limited ability to distinguish points in that space.

in perceptual psychology/psychophysics, there's the concept of the "just-noticeable difference" (JND) which is the smallest change to a stimulus you can make that is reliable detectable.

normally the JND is measured on physical properties like brightness, pitch, etc but there's no reason it couldn't be applied to a more abstract latent space. two points in a particular latent space may be mathematically unique, but if they're indistinguishable to humans we shouldn't treat them as distinct voices

	▲	altcunn 3 hours ago \| parent [-]
		[dead]