ASL-TFLite-Edge

This repository contains a TensorFlow Lite model trained to recognize American Sign Language (ASL) fingerspelling gestures using hand landmark data. The model is optimized for real-time inference on edge devices.

🧠 Model Details

• Format: TensorFlow Lite (.tflite)
• Input: 64x64 RGB image (generated from hand landmarks via Mediapipe)
• Output: Softmax probabilities over 59 ASL character classes (including a padding token)
• Frameworks: TensorFlow, Mediapipe

📁 Files Included

• asl_model.tflite – The TFLite model file for ASL recognition
• inference_args.json – JSON file containing the selected columns used for inference from parquet data
• tflite_inference.py – Inference script to run predictions from raw .parquet landmark files

🚀 How to Run Inference

You can download and load the TFLite model directly from Hugging Face using the huggingface_hub library.

Clone the image

git lfs install
git clone https://huggingface.co/ColdSlim/ASL-TFLite-Edge
cd ASL-TFLite-Edge

Requirements

pip install -r requirements.txt

Run the Script

python tflite_inference.py path/to/sample.parquet

Output

Predicted class index: 5

🔁 You can map this class index back to the character using your char_to_num mapping used during training.

📌 Example Workflow

1. Extract right-hand landmark data from Mediapipe and store it in a .parquet file.

2. Ensure it contains the same selected_columns as in inference_args.json.

3. Run tflite_inference.py to get the predicted class.

🧾 License

This project is licensed under the Apache 2.0 License.

👨‍💻 Author

Developed by Manik Sheokand

For sign language fingerspelling Recognition on edge devices using TensorFlow Lite