[TYPE_FASTER]

Each segment of the trailer is a different MP4. They are all hosted on the AWS CloudFront CDN. My guess is, like somebody stated, they pre-rendered many videos, and use a lambda at the edge to decode a GUID in the URL into a S3 path.

[remram]

Those are probably not GUIDs but hashes used for cache busting. Changing them doesn't change the content. Running a lambda for this also seems uselessly costly.

[echelon]

They probably used concatenative audio rendering since the digits are tractable to get the actor to record. (Probably just on the borderline in terms of effort for now.) But can you imagine when they are able to use real time text to speech with high fidelity?

We're not far from emotive TTS + photo realistic rendering in a real time engine, potentially with neurally generated models to match real time events or camera input.

Matrix 5 will be this, but you'll be in the trailer, and it'll reference real world events. And that's just the start.

The future of audio and video is going to be crazy.

[jjoonathan]

What does the state of the art look like? I'm not following closely enough to know, I only get an update when something leaks into the meme-o-sphere, like the recent "World Leaders singing Numa Numa:" https://www.youtube.com/watch?v=Y8CKlTlQA64

[echelon]

Here's my stupid text to speech website:

https://vo.codes

And you can try the alpha of "version 2" by visiting this, which has improvements in synthesis and vocoder quality, and also lets people contribute:

https://api.vo.codes/enable

This is far from state of the art, but you can see at the other end of the spectrum that it's an easy toy that anybody can use.

I'm working on a photo -> 3D rigged model system and model style transfer. You can follow me on Twitch or Twitter to see that when it's ready.

[jjoonathan]

Very cool! This is fun and I can see big improvements from v1 to v2. I look forward to watching this evolve!

https://vo.codes/tts/result/TR:s1rj02g34ppc7bhq1m8bf6p3thny6

[echelon]

Thanks!!

A few areas for improvement in the clip you posted:

I need to add better duration estimation. It's unfortunately truncated.

A lot of the community-trained voices don't fully leverage phonetic annotation, so some of the words fall flat.

I think the synthesizer has too much noise in it (you can see this in the image). The person who trained it probably used noisy data.

Finally, the universal vocoder isn't handling James Earl Jones' deep voice very well. It should be fine tuned.

[EGreg]

The future of audo and video is deepfakes for everything. Won’t be able to trust any video evidence anymore — except if watermarked by a trusted signatory.

[iolo]

I wouldn't even trust a watermark. It'd have to be a cryptographic signiture on the file itself.