Chaplin-UI is a gentle, privacy-focused tool that reads lips and turns them into text. Simply record yourself speaking (or upload a video), and watch as your words appear on screen—all without making a sound. This project is based on Chaplin by Amanvir Parhar, with added web interface and UI improvements. The VSR model achieves 19.1% word error rate (WER) on LRS3. Perfect for:

• 🎤 Silent communication - Type without speaking, or transcribe existing videos
• 🔒 Privacy-first - Everything runs locally on your machine (Privacy Policy)
• 🌐 Web-based - Works in any modern browser—no installation needed
• 🎨 Beautiful UI - Clean, calming design that adapts to your system theme

I built this after a week of laryngitis—when I couldn't speak, I needed a way to communicate. If you've ever wanted to say something without making a sound, Chaplin-UI might help:

• Public places — Libraries, offices, late-night calls, or anywhere you want to stay quiet
• Deaf and hard-of-hearing — Mouth words to communicate when sign language isn't shared
• Medical conditions — ALS, aphonia, cerebral palsy, laryngectomy, Parkinson's, vocal cord paralysis, selective mutism
• Temporary voice loss — Laryngitis, recovery from throat surgery, or vocal strain
• Privacy — Situations where you'd rather not speak aloud but still need to get your words out

Apple just acquired a silent-speech company (Q.ai) for $2 billion—this space matters, and open-source tools like this keep the technology accessible.

• Python 3.12+ (check with python3 --version)
• LLM local server – Ollama or LM Studio (the app finds one you have running):
  • Ollama – ollama serve + ollama pull <model>
  • LM Studio – load model, enable Local Server (port 1234)
• Modern web browser with camera access

1. Clone the repository:

   git clone https://github.com/loganngarcia/chaplin-ui.git
   cd chaplin-ui

2. Set up Python environment:

   python3 -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   pip install -r requirements.txt

3. Download model files:

   ./setup.sh

   This downloads the VSR model from Hugging Face (~500MB).

4. Start your LLM server (pick one: Ollama or LM Studio):

   Option A – Ollama

   ollama serve        # usually starts automatically
   ollama pull llama3.2   # or mistral, llama2, etc.

   Option B – LM Studio

   • Open LM Studio
   • Load a model (we recommend zai-org/glm-4.6v-flash)
   • Go to Developer tab → Enable Local Server (port 1234)

5. Run the web app:

   ./run_web.sh

   The UI opens in ~1 second at http://localhost:8000. The model loads in the background (~30–60 sec first time); you can use the interface right away—buttons enable when ready.

6. Start transcribing:

   • Record live: Click "Start recording" to capture video from your camera
   • Upload a video: Click "Upload video" to transcribe an existing video file
   • Your transcription appears in both raw and corrected formats. (You can change the LLM in settings if needed.)

   That's it! The app handles everything else for you.

chaplin-ui/
├── chaplin_ui/              # Core shared modules
│   └── core/                # Shared utilities, models, configs
│       ├── models.py        # Pydantic data models
│       ├── constants.py     # All configuration constants
│       ├── llm_client.py    # LLM API wrapper
│       ├── video_processor.py # Video processing utilities
│       └── ...
├── web/                     # Web app frontend
│   ├── index.html          # Main HTML
│   ├── style.css           # Styles (Apple HIG)
│   └── app.js              # Frontend logic
├── web_app.py              # FastAPI backend server
├── chaplin.py              # CLI implementation
├── main.py                 # CLI entry point
└── pipelines/              # VSR model pipeline

Simply put, Chaplin-UI watches how your lips move and turns that into text. Here's what happens behind the scenes:

1. You provide video - Either record yourself speaking or upload an existing video file
2. Face detection - The app finds and tracks your face in the video
3. Lip reading - A trained model watches your lip movements and creates initial text
4. Text refinement - An AI language model cleans up the text, adds punctuation, and fixes any mistakes
5. You see results - Both the raw transcription and the polished version appear on screen

You can copy the corrected text with one click, or review the raw output to see what the lip-reading model detected.

Chaplin-UI supports two local LLM backends. Both use OpenAI-compatible APIs:

• Web app: Select provider in the "LLM Provider" dropdown and optionally override the model name.
• CLI: Use llm_provider=ollama or llm_provider=lmstudio, or run with --config-name ollama for Ollama defaults.

• chaplin_ui/core/ - Shared code used by CLI and Web interfaces
• web_app.py - FastAPI server handling video uploads and processing
• chaplin.py - CLI version with keyboard typing
• pipelines/ - VSR model inference pipeline

Web App:

source .venv/bin/activate
python web_app.py

CLI:

source .venv/bin/activate
python main.py config_filename=./configs/LRS3_V_WER19.1.ini detector=mediapipe
# With Ollama:
python main.py --config-name ollama
# Or: python main.py llm_provider=ollama llm_model=mistral

We follow Python best practices:

• Type hints on all functions
• Docstrings (Google style) for all public functions
• Logging instead of print statements
• Constants centralized in chaplin_ui/core/constants.py

# Test imports
python -c "from chaplin_ui.core import *; print('✓ All imports work')"

# Test web app
python web_app.py &
curl http://localhost:8000/api/health

We love contributions! Whether it's:

• 🐛 Bug fixes
• ✨ New features
• 📝 Documentation improvements
• 🎨 UI/UX enhancements
• 🔧 Code refactoring

See our Contributing Guide for details on:

• How to set up your development environment
• Code style guidelines
• How to submit pull requests
• Where to ask questions

First time contributing? Check out our good first issues!

This project is licensed under the MIT License - see LICENSE for details.

Chaplin-UI is based on Chaplin by Amanvir Parhar. We're grateful for the original work that made this project possible!

• VSR Model: Based on Auto-AVSR by mpc001 (19.1% WER on LRS3)
• Dataset: Lip Reading Sentences 3
• LLM: Uses Ollama or LM Studio for local text correction (both OpenAI-compatible)

• 🐛 Found a bug? Open an issue
• 💡 Have an idea? Start a discussion
• 📧 Questions? Check our FAQ or open a discussion