So, the multi-speaker neural TTS models have gotten eerily good; the few-sample/voice-cloning NNs are also now credible with a few minutes of samples. Has anyone told the various media fandoms they could get their favorite characters' voices on tap with some elbow-grease...?
-
Show this thread
-
Replying to @gwern
Without an implementation that also does style transfer for emotion and inflection, it's hard to make media using these tools. But yeah, technically we're a couple of thesis projects away from being able to resurrect any show as a fan cartoon with the original voice actors.
1 reply 0 retweets 7 likes -
Replying to @AndreTI
My thinking is normal conversational speech is already a big chunk of what people would use it for. I also suspect existing models may encode substantial emotion/inflection range into the embeddings/latents already which can be reverse-engineered & then controlled like w/GANs.
2 replies 0 retweets 6 likes -
"$CHARACTER narrates your voicemail message" or "$CHARACTER narrates your Facebook comment" would be extremely popular product features, concerns about aesthetics regarding artistic integrity or opinion of rightsholders aside.
2 replies 1 retweet 7 likes -
Checking, voices can't be copyrighted... but voice cloning used in conjunction with art/names/music violates copyrights. More confusing is if just the voice with serial numbers filed off is a violation of privacy/'publicity rights' - many voices sound nothing like the voice actor
1 reply 0 retweets 3 likes -
Yeah. There are also potentially issues (or at least rightsholders would probably assert issues) with trade dress, obstruction of business, etc etc.
1 reply 0 retweets 1 like -
(I'll admit a little bit of sympathy for their arguments here. The reason folks would want this is to have the convincing simulcrum of Mickey Mouse say things that Mickey has never said, and Disney will be *extraordinarily* interested in having Mickey continuing to not say things
1 reply 0 retweets 2 likes
And there's a pretty interesting argument on e.g. what the difference of an impersonator doing an impression of a character is as an artistic work versus a (indistinguishable?) simulacrum of the character is. Is that *even more* transformative, or less so? Is it *too* good?
-
-
I'm wincing thinking about fueling the 'derivative vs transformative use of dataset' argument that is *still* unresolved for deep learning & ML in general - does a trained model, and its outputs, constitute a new work, or is it derivative of *all* copyrights anywhere in training?
0 replies 0 retweets 7 likesThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.