Hacker News new | ask | show | jobs
by lyu07282 252 days ago
Yeah in the future we might use some sort of learned spatial+temporal representation to compress video, same for audio. Its easier to imagine for audio: Instead of storing the audio samples, we store text + some feature vectors that uses some model to "render" the audio samples.
1 comments
whycome 252 days ago
It’s not absurd to think that you could send a model of your voice to a receiving party and then have your audio call just essentially be encoded text that gets thrown through the voice generator on the local machine.

AI video could mean that essential elements are preserved (actors?) but other elements are generated locally. Hell, digital doubles for actors could also mean only their movements are transmitted. Essentially just sending the mo-cap data. The future is gonna be weird

link
JimDabell 252 days ago
Yeah, I brought that up here and got some interesting responses:

> It would be interesting to see how far you could get using deepfakes as a method for video call compression.

> Train a model locally ahead of time and upload it to a server, then whenever you have a call scheduled the model is downloaded in advance by the other participants.

> Now, instead of having to send video data, you only have to send a representation of the facial movements so that the recipients can render it on their end. When the tech is a little further along, it should be possible to get good quality video using only a fraction of the bandwidth.

— https://news.ycombinator.com/item?id=22907718

Specifically for voice, this was mentioned:

> A Real-Time Wideband Neural Vocoder at 1.6 Kb/S Using LPCNet

— https://news.ycombinator.com/item?id=19520194

link
whycome 252 days ago
In the future, our phone contacts will store name, address, phone number, voice model. (The messed up part will be that the user doesn’t necessarily send their model, but the model could be crafted from previous calls)

You could probably also transmit a low res grayscale version of the video to “map” any local reproduction to. Kinda like how a low resolution image could be reasonably reproduced if an artist knew who the subject was.

link