SignCoach

SignCoach is a Tkinter desktop app for practicing the 24 static ASL alphabet signs. It includes:

• A Learn tab with reference drawings and live camera feedback.
• A Quiz tab with 10 random letter challenges.
• Optional MediaPipe hand cropping.
• Optional TensorFlow MobileNetV2 training and saved-model loading.

Run

pip install -r requirements.txt
python signcoach.py

On first launch, no model is trained. Use File > Train model... or load an existing .keras model and matching .meta.json file.

Train From WSL/CUDA

source ~/mathenv/bin/activate
cd /mnt/c/Users/Alex/Documents/Codex/2026-06-04/files-mentioned-by-the-user-pasted/outputs

python signcoach.py train \
  --data "/mnt/c/Users/Alex/Documents/Codex/2026-06-04/files-mentioned-by-the-user-pasted/outputs/training_data/asl_dataset" \
  --out signcoach_model \
  --epochs 15 \
  --batch-size 32

The CLI training command prints batch and epoch progress, then saves:

• signcoach_model.keras
• signcoach_model.meta.json
• signcoach_model_history.csv
• signcoach_model_learning_curve.png when matplotlib is installed

Dataset

Use a dataset folder with one subfolder per letter, such as A, B, C, etc. The app excludes J and Z because those are motion signs.

Recommended dataset: Kaggle ASL Alphabet dataset.