License clarification: Are melspectrogram.onnx and embedding_model.onnx under Apache 2.0 and suitable for commercial use?

[RellyLee]

Hi, thanks for the great work on openWakeWord!

I'd like to clarify the licensing of the two shared feature extraction models: melspectrogram.onnx and embedding_model.onnx.

The README states that:

All code in the repository is licensed under Apache 2.0.
All included pre-trained models are licensed under CC BY-NC-SA 4.0 due to training data with unknown or restrictive licensing.
However, I understand that melspectrogram.onnx and embedding_model.onnx are distinct from the wake word classification models — specifically:

melspectrogram.onnx was exported from a torchlibrosa-based PyTorch implementation (as shown in notebooks/converting_google_speech_embedding_model.ipynb), which is Apache 2.0 code.
embedding_model.onnx is a manual Keras re-implementation of Google's speech_embedding model, which Google releases under Apache 2.0. The README also explicitly notes this (under Model Architecture).
My questions are:

Are melspectrogram.onnx and embedding_model.onnx covered by the repository's Apache 2.0 license (not CC BY-NC-SA 4.0), given their origin from Apache 2.0 sources?
Is it permitted to use these two models in a commercial product, as long as we train our own wake word classification model on top of them (i.e., we do not use the pre-trained alexa, hey_jarvis, etc. classification models)?
Any clarification would be greatly appreciated. Thank you!

[jasonkellem]

+1 on this — I'm in the same position: building a commercial accessibility app, training our own classifier, and needing certainty on just these two files. The reasoning above matches mine: the README's stated rationale for CC BY-NC-SA is training data with restrictive licensing, and neither of these files was trained on that data — melspectrogram is a fixed DSP transform, and the embedding is Google's Apache-2.0 module re-implemented.

@dscripka — even a one-line confirmation would unblock a fair number of downstream projects. If it helps, the README's offer to add models with more permissive licensing on request suggests this is the intent already; making it explicit for the two shared feature models would settle it.