parakeet-redux-c

Parakeet Redux weights for parakeet-redux-c

parakeet-redux.bin is the single weights file read by parakeet-redux-c, a dependency-free C implementation of moondream/parakeet-redux, the 1.58-bit version of NVIDIA's parakeet-tdt-0.6b-v3.

It is converted by convert.py from the ONNX export at eschmidbauer/parakeet-redux-onnx without recomputing anything: the ternary encoder weights are stored at 2 bits per weight with their float16 group scales, everything else is the ONNX tensor under a stable name, and vocab.txt and the decoding settings are embedded. The file is architecture independent and about 250 MB; the C runtime packs it into its kernel layouts while loading.

Use

git clone https://github.com/eschmidbauer/parakeet-redux-c
cd parakeet-redux-c
git clone https://huggingface.co/eschmidbauer/parakeet-redux-c models/parakeet-redux-c
make
./build/parakeet speech.wav

Convert it yourself

pip install -r requirements.txt
python convert.py                                  # downloads the ONNX export
python convert.py --onnx-dir path/to/parakeet-redux-onnx

The file layout is documented at the top of convert.py: a 256-byte header with the decoding settings, a table of named tensors, then the data 64-byte aligned. Ternary codes are row major, four weights per byte, value = code - 1.

License and attribution

Converted from moondream's parakeet-redux, itself derived from NVIDIA's parakeet-tdt-0.6b-v3; both are released under CC BY 4.0, and so is this file. Neither moondream nor NVIDIA endorse this conversion.