moonshine-c

A zero-dependency C implementation of Moonshine, a fast and accurate automatic speech recognition (ASR) model.

All weights are exported as float32 from HuggingFace.

Getting Started

1. Install Python dependencies

pip install -r scripts/requirements.txt

2. Export float32 model weights

Downloads weights from HuggingFace and converts them to .bin files.

# Export both tiny and base (default)
python scripts/export-weights.py

# Export only tiny
python scripts/export-weights.py --model tiny

Output goes to models/<model>/ (e.g. models/tiny/, models/base/), each containing:

• encoder.bin — float32 encoder weights
• decoder.bin — float32 decoder weights
• tokenizer.bin — BPE tokenizer (downloaded and converted from HuggingFace)

3. Build and run

make
./test_process models/tiny jfk.wav