openWakeWord featurization graphs (mirror)

This repo mirrors the two ONNX feature-extractor models that openWakeWord uses as shared frontends for every wake-word DNN it trains:

File	Size	Purpose
`melspectrogram.onnx`	1.1 MB	Converts 16 kHz int16 audio → 32-bin mel-spectrogram frames
`embedding_model.onnx`	1.3 MB	Google `speech_embedding/1` — mel frames → 96-dim embeddings

These are NOT trained by Little Bear Labs — they are upstream Apache-2.0 assets from the openWakeWord project, mirrored here so HuggingFace hf_hub_download callers (e.g. witness-wake) can pull them without pip / GitHub-release-API round trips.

When to use this

Use this mirror if you're:

Building a Rust wake-word runtime that uses ort + the openwakeword feature pipeline directly (no Python).
In an environment where the openwakeword pip package isn't available.
Wanting a fixed HuggingFace URL with hf_hub_download semantics.

Otherwise, prefer:

pip install openwakeword then python -c "import openwakeword; openwakeword.utils.download_models([])" — populates <site-packages>/openwakeword/resources/models/.
The original openWakeWord GitHub release assets.

Pairing with a wake-word DNN

These two graphs are the shared frontend. Pair them with any openwakeword DNN, e.g.:

littlebearlabs/hey-virgil-wake-word — Little Bear Labs's "hey virgil" detector.
Any community-trained wake-word .onnx from the openwakeword ecosystem.

License

Apache-2.0 (inherited from openWakeWord + Google speech_embedding).

Citation

If you use this in research, cite openWakeWord:

@software{openwakeword,
  author = {David Scripka},
  title = {openWakeWord: A library for training open-source wake word models},
  year = {2024},
  url = {https://github.com/dscripka/openWakeWord}
}

Downloads last month: -; Downloads are not tracked for this model. How to track