Licence of the shared feature models (melspectrogram, embedding_model) for a commercial product

[rmunoz928]

Hi David — thank you for openWakeWord.

We are building a commercial voice assistant and intend to train our own wake-word model with your code, shipping none of the pre-trained ones. Three questions; a one-line answer to each would be enough:

1. Are melspectrogram.onnx / .tflite and embedding_model.onnx / .tflite covered by the repository's Apache 2.0 licence, or by the CC BY-NC-SA 4.0 that the README applies to the pre-trained models? The README describes the embedding model as a re-implementation of Google's Apache-2.0 speech_embedding, and the melspectrogram model as an ONNX implementation of Torch's melspectrogram function with fixed parameters — neither of which seems to involve the "datasets with unknown or restrictive licensing" the README gives as the reason for the NonCommercial terms.
2. If they are NonCommercial, is there a commercial-use path you would point us to?
3. For the avoidance of doubt: may a classifier we train ourselves, on data we license ourselves, that loads those two feature models, be used commercially?

This repeats #313, #338 and #346 — very happy for an answer in any of them rather than here. We will attribute openWakeWord prominently either way.