Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

1.8 KB
Raw Blame

X post: the Audio2Face forum write-up

Announcement copy for the NVIDIA Developer Forums post that went live on 2026-09-22: https://forums.developer.nvidia.com/t/a-digital-human-in-a-browser-tab-streaming-audio2face-3d-onto-whatever-rig-the-visitor-brought/383953

Owner posts by hand. Every number below is the one measured against production in the forum post itself. Attach a short screen recording of https://three.ws/demos/audio2face if posting with media; X rewards a clip of a face moving far more than a link card.

Option A, the demo first (recommended)

Type a sentence, get a 3D character. Then it talks to you, with lips driven by NVIDIA Audio2Face-3D. In a browser tab, no engine, no install.

The hard part was never the model. It was the rig: the visitor's avatar might speak ARKit, VRM vowels, or Oculus visemes. We drive all three from one track.

Wrote up how, with latencies measured on prod: [link]

Option B, the number

Audio2Face-3D animates a 5.5 s line of speech in 1.3 to 2.2 s, in a browser, on a rig it has never seen.

How we map an ARKit-52 track onto VRM and Oculus rigs, why you must play the 44.1 kHz audio and not the 16 kHz copy you fed the model, and the function id that died silently: [link]

Option C, the pair

Two posts on the NVIDIA Developer Forums this week.

One about a lane that broke: three NIM models retired, a 410 that a fallback chain hides for two weeks, and the fixes.

One about a lane that shines: Audio2Face-3D driving a WebGL avatar the visitor generated ninety seconds ago.

Both with the real code and the real numbers. [audio2face link] [retirement link]

Reply to pin under any of them

Try it, no signup: https://three.ws/demos/audio2face Source: github.com/nirholas/three.ws (api/_lib/a2f-nvidia.js, src/voice/a2f-player.js)