Talking Head

A 2D female talking head that reads text aloud in English or Hebrew, with lip sync. Static site, no build step, no API keys.

Run

• PC: double-click index.html, or serve the folder: python -m http.server 8000 → http://localhost:8000
• Phone (same Wi‑Fi): python -m http.server 8000 --bind 0.0.0.0, then open http://<your-PC-IP>:8000 on the phone.
• Anywhere: upload the folder to any static host (GitHub Pages, Netlify, Cloudflare Pages).

How it works

Part	File	Notes
Speech	js/app.js	Browser Web Speech API (speechSynthesis). Picks a female voice by name; changeable under Voice settings.
Visemes	js/visemes.js	Converts spelling to 12 mouth shapes (English rules; Hebrew letters + niqqud, guessing missing vowels).
Sync	js/app.js	Snaps to the engine's word-boundary events when available; otherwise a timer whose speed self-calibrates from past utterances.
Avatar	js/avatar.js	Original SVG character with a parametric mouth (smoothly blended), blinking, gaze and head motion.
Updates	js/update-check.js	Every 5 min and on returning to the tab, compares the deployed files with the loaded ones and offers a reload. Add new files to its FILES list.
Looks	js/looks.js	8 selectable cartoon characters (hair, colours, accessories) sharing one head rig, plus photo avatars. Add a look by adding an entry. Male looks prefer male voices.
3D avatar	js/avatar3d.js	"Mia 3D": a GLB character rendered with TalkingHead 1.7.0 and three.js 0.180.0 (from jsDelivr, via the import map in index.html). Our viseme timeline drives the model's Oculus viseme blend shapes, so the mouth, teeth and tongue are real 3D geometry. Loaded only when chosen.
Photo avatars	js/photo-avatar.js	Warps a portrait photo with a WebGL mesh on MediaPipe face landmarks (mouth, jaw, blinks, gaze, head sway) and paints teeth/mouth interior. "Your photo" runs face detection in the browser; the photo never leaves the device.

Adding a bundled photo

1. Put a front-facing portrait (closed mouth, even light, ~768px) in photos/. Use AI-generated faces or photos of people who agreed.
2. Add { id, name, gender, photo: 'photos/x.jpg', landmarks: 'photos/x.json' } to js/looks.js.
3. The landmarks file is optional (without it, landmarks are detected on first use), but it saves phones a ~4 MB model download. To make one, pick the photo via "Your photo" and run JSON.stringify(AvatarLooks.find(l => l.upload).detected) in the browser console.

avatars/brunette.glb is the example avatar from the TalkingHead project, created at Ready Player Me and licensed CC BY-NC 4.0 (non-commercial use, with attribution, which the app shows on the stage). Other 3D avatars need a Mixamo-compatible rig plus ARKit and Oculus viseme blend shapes; check that their licence allows publishing on a public website.

photos/lena.jpg is a public-domain StyleGAN image from Wikimedia Commons (File:Woman 1.jpg).

Voices

Available voices depend on the device/browser:

• Windows: Microsoft Edge has natural online voices (English "Aria"/"Jenny", Hebrew "Hila"). Chrome only has the installed Windows voices; add Hebrew under Settings → Time & language → Speech.
• Android: Google Text-to-Speech, install the Hebrew voice data.
• iPhone/iPad: Hebrew "Carmit" and English "Samantha" are built in.