Imagine some day a hacker will be able to mimics speech with the next version of Tacotron 2[1] then use it over the next version of Face2Face... Scary future ahead.

[1] https://news.ycombinator.com/item?id=16014047

Wow those text to speech examples are amazing. The use in videogames could be huge. Exciting stuff.

My thoughts exactly!

On an earlier Tacotron article: https://news.ycombinator.com/item?id=15963888

I knew this was coming and it's more surreal that I thought it would be. We crossed the uncanny valley. I'm not sure we are prepared for trusted sources to be faked this well and this easily. It will be interesting to see how we as a society start to cope with an acid trip level of real life unreliable narrator.

This could be a new use for DRMs. Certifying that the image stream is real and unaltered (or only by trusted sources).

DRM does not do that. Plain old public key cryptography does.

I read something recently about photographers pushing for cryptographic keys, maybe it was even registering media in the blockchain. Certifying origin and authorship for exactly this situation. I think it was a podcast. Maybe Open Source.

Could be huge in dubbing or even just ADR for movies. I'll take a slightly uncanny smile over the abomination of some dubs any day.

Since this is around 18 months old, does anyone know what's the current state of the art?

I found this from ~3 months ago, the recent addition to github:

https://google.github.io/tacotron/publications/uncovering_la...

Seems to do with manipulating the speed styles in how we speak.

Listen to the "neutral" blue lagoon, vs token 6 (tired sound), vs mix w/ token 7 (prominence)" (very confident).

Paper: https://arxiv.org/pdf/1711.00520.pdf

Previous discussion: https://news.ycombinator.com/item?id=11314931 (2 years ago, 35 comments)

Isn't this around 18 months old now?

GANs should be able to do a lot better now.

Yes, this from 2016

Mods, please add 2016 to post title.