Multi-language sign language recognition system using neural networks, MediaPipe landmarks, and GPU optimization.

This project recognizes sign language across two alphabet sets:

• ASL (American Sign Language)
• ISL (Indian Sign Language)

SignLanguageProject_Pro/
├── data/
│   ├── raw/              # Original datasets
│   │   ├── ASL/          # Kaggle ASL Alphabet
│   │   └── ISL/          # Kaggle ISL Alphabet
│   └── processed/        # Extracted landmarks
│
├── models/               # Saved model weights
│
├── preprocess.py         # Image -> landmarks conversion
├── train.py              # GPU-optimized training
├── main.py               # Live recognition app
├── check_npy.py          # Quick NPY inspection utility
├── requirements.txt
└── README.md

pip install -r requirements.txt

python preprocess.py

python train.py

python main.py

Figure 1: The MLSLRS OpenCV local interface performing real-time inference on sign sequences.

Figure 2: The web-migrated deployment running in the browser with the glassmorphic UI.

• Framework: TensorFlow 2.15+
• Input: Hand landmarks (MediaPipe)
• Output: Sign prediction with confidence scores

Edit the following files to customize:

• train.py - Model architecture and training parameters
• preprocess.py - Dataset paths and preprocessing options
• main.py - Live app settings and confidence thresholds

• Landmarks are extracted using MediaPipe Hand Landmarker
• Temporal smoothing reduces jitter in sequences
• Models saved in Keras 2026 format (.keras)

Custom Project

Dhanush Alla