---

Highlights

• One command, nine sprites — feed in a portrait, get back 9 mouth shapes based on the Preston Blair phoneme system
• Auto face detection — MediaPipe finds the face automatically; no manual cropping needed
• AI-generated — Gemini (via OpenRouter) generates each mouth shape from reference artwork
• Production-ready output — transparent backgrounds, soft-edge feathering, WebP format
• Real-time lip sync engine — vanilla TypeScript engine syncs audio energy + text phonemes to shapes at 60fps
• Claude Code plugin — install the skill and let Claude run the full pipeline for you
• Zero frameworks — Python stdlib for HTTP, vanilla TS for the demo — minimal dependencies

How It Works

Portrait → Face Detection → Gemini AI → 9 Mouth Shapes → Post-Processing → Sprites
 (photo)    (MediaPipe)     (OpenRouter)  (Preston Blair)   (rembg + soft edges)   (WebP)

The 9 Shapes

Shape	Phonemes	Visual
A	M, B, P	Lips pressed together
B	S, T, D, K, N	Teeth visible, slightly open
C	A, I	Open mouth (speaking vowels)
D	E	Wide horizontal stretch
E	O	Rounded mouth
F	U, W, Q	Puckered lips
G	F, V	Upper teeth on lower lip
H	L, T, D	Barely parted lips
X	(idle)	Resting / closed mouth

---

Quick Start

Install

# Clone
git clone https://github.com/historacle-ai/blabberer.git
cd blabberer

# Install (pick one)
uv pip install -e .
# or: pip install -e .

Generate sprites

export OPENROUTER_API_KEY=sk-or-...   # https://openrouter.ai/keys

blabberer generate photo.jpg --output-dir ./sprites --verbose

Output:

sprites/
├── soft/              # Production-ready (transparent bg + soft edges)
│   ├── shape_A.webp
│   ├── shape_B.webp
│   ├── ...
│   └── shape_X.webp
├── shape_A.webp       # Transparent background (no soft edges)
├── ...
└── shape_X.webp

---

CLI Reference

blabberer generate <image> [options]

Options:
  --api-key KEY          OpenRouter API key (or set OPENROUTER_API_KEY env var)
  --output-dir, -o DIR   Output directory (default: ./sprites)
  --crop-box L,T,R,B     Manual face crop box (auto-detects if omitted)
  --model MODEL          Gemini model (default: google/gemini-3.1-flash-image-preview)
  --workers N            Concurrent API calls (default: 3)
  --skip-postprocess     Skip background removal and soft edges
  --verbose, -v          Detailed output

Python API

from blabberer import generate_sprites, full_postprocess

# Generate sprites
results = generate_sprites(
    image_path="photo.jpg",
    output_dir="./sprites",
    api_key="sk-or-...",
)

# Post-process (background removal + soft edges)
final = full_postprocess("./sprites")

Face Detection

from blabberer import detect_face_crop
from PIL import Image

image = Image.open("photo.jpg")
crop_box = detect_face_crop(image)  # → (left, top, right, bottom)

# Custom expansion ratios
crop_box = detect_face_crop(
    image,
    expand_top=0.3,      # 30% above face
    expand_bottom=0.5,   # 50% below face
    expand_sides=0.25,   # 25% each side
)

---

Interactive Demo

A browser demo lets you see lip sync in action — audio analysis drives mouth shapes in real time.

cd demo && npm install && npm run dev
# → http://localhost:5173

Two modes:

Mode	Entry	What it does
Canvas (default)	simple.html	Draws shapes on a <canvas> overlay — no sprites needed
Sprite	index.html	Dual-layer crossfade with generated sprite images

Lip Sync Algorithm

The LipSyncEngine uses a 3-tier energy model with text-informed shape selection:

1. Audio analysis — Web Audio AnalyserNode extracts frequency data (first 30 bins)
2. Energy smoothing — exponential moving average (rise=0.28, decay=0.18)
3. Hysteresis state machine — LOW / MID / HIGH with asymmetric thresholds to prevent flicker
4. Text-to-shape mapping — 28 letter + 12 digraph preferences per energy level
5. Min hold time — 100ms per shape prevents "chewing" artifacts
6. Crossfade rendering — two stacked <img> elements with 60ms CSS transitions

Use the Engine in Your Project

The demo's src/ files are standalone vanilla TypeScript — zero framework dependencies. Copy them into any project:

import { AudioAnalyser } from './audioAnalyser';
import { LipSyncEngine } from './lipSync';

const audio = document.querySelector('audio')!;
const analyser = new AudioAnalyser();
const node = analyser.connect(audio);
const engine = new LipSyncEngine(node!);

engine.loadText('Hello world');
audio.play();

function animate() {
  requestAnimationFrame(animate);
  const shape = engine.update(performance.now());
  // shape: 'A' | 'B' | 'C' | 'D' | 'E' | 'F' | 'G' | 'H' | 'X'
}
requestAnimationFrame(animate);

---

Claude Code Plugin

Blabberer ships as a Claude Code plugin with a /lip-sync-pipeline skill that automates the entire workflow.

Install

/plugin install historacle-ai/blabberer

Run

/blabberer:lip-sync-pipeline photo.jpg

Claude handles face detection, sprite generation, post-processing, demo setup, and troubleshooting — end to end.

---

Requirements

Component	Requirement
Python	>= 3.10
API key	OpenRouter with Gemini model access
Disk	~170 MB for rembg U2Net model (downloaded on first run)
Node.js	>= 18 (demo only)

Project Structure

blabberer/
├── blabberer/                # Python package
│   ├── cli.py                # CLI entry point
│   ├── generator.py          # Face detect → Gemini → sprites
│   ├── face_detect.py        # MediaPipe face detection
│   ├── shapes.py             # 9 Preston Blair shape definitions
│   ├── postprocess.py        # Background removal + soft edges
│   └── reference/            # 12 bundled reference images
├── demo/                     # Browser lip sync demo
│   ├── simple.html           # Canvas mode (default)
│   ├── index.html            # Sprite mode
│   └── src/                  # Vanilla TypeScript
│       ├── lipSync.ts        # Audio + text → shape engine
│       ├── audioAnalyser.ts  # Web Audio API bridge
│       ├── mouthRenderer.ts  # Dual-layer crossfade
│       ├── simple.ts         # Canvas demo
│       └── main.ts           # Sprite demo
├── skills/                   # Claude Code plugin
│   └── lip-sync-pipeline/
├── .claude-plugin/           # Plugin manifest
└── pyproject.toml

---

Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request.

1. Fork the repo
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

---

About HistOracle

Blabberer was extracted from the production lip sync pipeline at HistOracle — an AI platform that brings historical figures to life through interactive conversations. HistOracle uses AI-generated visemes, real-time audio analysis, and natural language understanding to create lifelike talking portraits of figures from history.

We open-sourced the lip sync sprite generation and animation engine so developers, educators, and creators can build their own talking-face experiences.

License

MIT License — see LICENSE.

Copyright (c) 2026 Editlingo Solutions LLC