NedAlastor/ASl_by_letter_translator_model

Real-time ASL LSTM model

Short description

• This repository contains Keras LSTM model artifacts for a real-time American Sign Language (ASL) alphabet classifier trained on hand-landmark sequences.

Files included

• lstm_model.h5 — main Keras model (HDF5) exported from training
• lstm_model_backup.h5 / lstm_model_backup_pre_augment.h5 — backup checkpoints
• lstm_model_label_encoder.pkl / label_encoder.npz — label encoder mapping class indices ↔ labels

How to use

• Download the files from the Hugging Face repo and load in Python with Keras:

from tensorflow import keras
model = keras.models.load_model('lstm_model.h5')

To decode predictions, load the label encoder that was used during training.

Training data

• The model was trained on landmarks extracted from video frames (see repository docs for the preprocessing pipeline). For details on the dataset composition and augmentation see the project repo.

Evaluation

• See repository training logs (rtslt/train_log.txt) and evaluation scripts in rtslt/ml_models/evaluate_models.py for metrics and confusion matrices.

Limitations and intended use

• Intended for research and prototyping. Do NOT rely on this model for safety-critical applications.
• Performance may drop on different camera angles, lighting, or unseen hand shapes; re-training or fine-tuning on target data is recommended.

Licensing and citation

• Add your preferred license here. If you want me to add a specific license file (e.g., MIT), tell me and I will add it.

Contact

• See the main repo for author and contact information.