LipReader

A lightweight, CPU-only lip reading toolkit for command recognition from video.
No GPU required — runs efficiently on Intel i5 and similar systems.

✨ Features

• CPU-only: No GPU or deep learning dependencies.
• CLI & API: Use via command line or import as a Python library.
• Trainable: Learn custom lip motion patterns from your own videos.
• JSON-based: All data stored in human-readable JSON format.
• Real-time ready: Optimized for low-latency inference.

📦 Installation

Install in development mode (recommended):

git clone https://github.com/Parhamfakhar1/lipreader.git
cd lipreader
pip install -e .

Requires: Python 3.7+, OpenCV, NumPy

🚀 Usage

Train a new command

lipreader train --video start.mp4 --word start

You can train the same word multiple times with different videos:

lipreader train -v start1.mp4 -w start
lipreader train -v start2.mp4 -w start

Predict from a video

lipreader predict --video test.mp4

Sample output:

🎯 Prediction: start

📈 Probabilities:
   start: 86.3%
   stop: 13.7%

CLI Options

💻 Python API

Use LipReader directly in your code:

from lipreader import LipReader

# Initialize
reader = LipReader("commands.json")

# Train
reader.train("start.mp4", "start")

# Predict
predicted_word, probabilities = reader.predict("unknown.mp4")
print(f"Detected: {predicted_word}")

🗃️ Data Format

All trained patterns are saved in lip_data.json:

{
  "start": {
    "samples": [
      {
        "avg_ratio": 1.28,
        "ratio_std": 0.25,
        "min_ratio": 0.78,
        "max_ratio": 1.88,
        "frame_count": 120,
        "video": "start1.mp4"
      }
    ]
  }
}

⚠️ Limitations

• Works best in good lighting with front-facing video.
• Accuracy depends on clear lip motion (silent articulation works).
• Not designed for full-sentence lip reading — optimized for short commands.

📄 License

MIT License