This project aims to develop a real-time recognition system for French Sign Language (LSF). It detects signs from video input, converts them into raw text, reformulates the text into natural French using NLP, and generates audio output with text-to-speech (TTS). Subtitles are displayed for accessibility. The goal is to bridge communication gaps, with plans to support multiple sign languages for global impact.

• Sign detection using MediaPipe Hands.
• Sign classification with a LSTM model (TensorFlow/Keras).
• Text reformulation with T5-small (Hugging Face Transformers).
• Audio synthesis using gTTS.
• Interactive web interface with Streamlit.
• Compatible with the LSF-Data dataset (parlr/lsf-data) or custom datasets.

• Python: Core backend.
• MediaPipe: Real-time hand tracking.
• TensorFlow/Keras: LSTM for sign classification.
• Hugging Face Transformers: T5-small for text reformulation.
• gTTS: Text-to-speech.
• Streamlit: Web interface.
• OpenCV: Video processing.
• Docker: Deployment.

Under active development. Initial setup includes repository structure and documentation. Next steps: dataset integration and model training.

1. Clone the repository:

   git clone https://github.com/fless-lab/lsf-recognition.git

2. Further setup instructions coming soon.

lsf-recognition/
├── data/               # Raw and processed datasets
├── models/             # Trained and pretrained models
├── src/                # Source code (detection, classification, NLP, UI)
├── notebooks/          # Data exploration notebooks
├── tests/              # Unit tests
├── Dockerfile          # Containerization
├── requirements.txt    # Python dependencies
├── README.md           # Project documentation
├── LICENSE             # MIT License

Contributions are welcome! Check the issues for open tasks or submit your ideas. Follow standard practices (tests, documentation, PEP8).

This project is licensed under the MIT License.

For questions, open an issue or reach out via GitHub.

Stay tuned for updates as the project grows!