Real-Time 3D Talking Avatar in the Browser — Three.js + ARKit Blendshapes + Audio Lip-Sync

I need a realistic 3D female avatar that runs live in Chrome inside a React web application, with audio-driven lip-sync and natural idle behaviour. This is not a render or animation job. No pre-rendered video, no Unity or Unreal builds, no cinematics. The final deliverable must run in a browser tab in real time at 30fps on an Apple Silicon Mac. Audio is generated on demand at runtime, so no animation can be pre-baked — lip-sync must be driven from an arbitrary audio stream. Deliverables 1. Avatar asset .glb (glTF 2.0), plus source files (.blend or Maya) and all textures ARKit 52 blendshapes, named exactly per Apple's specification — this is non-negotiable and drives the lip-sync 30,000–60,000 triangles PBR textures, 2K (4K acceptable for face only) Head-and-shoulders bust is fine; full body a plus Realistic adult female, professional appearance, neutral and friendly resting expression 2. React component React + Three.js (react-three-fiber acceptable) The open-source TalkingHead JS library is an acceptable and encouraged basis Must expose this API: js avatar.speak(audioUrl, text) // starts lip-synced playback avatar.stopSpeaking() avatar.setEmotion("neutral" | "happy" | "concerned") 3. Idle behaviour (this is what makes it feel alive — please do not skip) Irregular natural blinking Subtle breathing Micro head movements and weight shifts Smooth transitions between idle → speaking → idle 4. Performance 30fps minimum in Chrome on an Apple Silicon Mac Total asset payload under 15MB No main-thread blocking while idle Licensing — please read before bidding The avatar must be fully licensed for use and redistribution inside a commercial interactive web application. Models derived from Daz3D, Reallusion / Character Creator, or similar bases frequently prohibit this. If your workflow uses any such base, disclose it in your proposal. Undisclosed licensing issues discovered later will void the contract. Full rights transfer on final payment — source files, textures, and component code. Milestones (20 / 30 / 30 / 20) 20% — Concept: three face/style options as renders, plus written confirmation of the licensing basis 30% — Modelled and textured head delivered as .glb, blendshapes verified by name 30% — Working React component with lip-sync, delivered as a live demo URL 20% — Idle behaviour complete, performance targets met, source and rights transferred Acceptance criteria .glb loads in a standard glTF viewer with all 52 blendshapes present and correctly named I can call speak() with my own arbitrary audio file and get accurate lip-sync Runs at 30fps+ in Chrome on Apple Silicon Idle animation runs indefinitely with no visible looping artifacts No console errors To apply, answer these four questions Send a live URL where I can see your previous avatar or character work running in a browser. Video reels will not be reviewed. What is your approach to audio-driven lip-sync in the browser? Have you worked with the ARKit 52 blendshape set before? What is the licensing basis of the models you build from? Proposals without a live URL will not be considered.