Third-party licenses
The wake word classifier bundled with this application was trained with the help of third-party models and datasets. This directory collects the corresponding license texts and attribution notes.
Components
| Component | Source | License | License text |
|---|---|---|---|
Silero VAD (models/silero_vad_16k_op15.onnx) |
https://github.com/snakers4/silero-vad | MIT | models/LICENSE |
openWakeWord feature extractor (models/openwakeword_melspectrogram.onnx, models/openwakeword_embedding_model.onnx) |
https://github.com/dscripka/openWakeWord | Apache-2.0 | LICENSE.openwakeword.txt |
| MUSAN corpus (noise negatives used for training) | http://www.openslr.org/17 | CC BY 4.0 | LICENSE.musan.txt |
| FUSS dataset (noise negatives used for training) | https://github.com/google-research/sound-separation/tree/master/dcase2020_fuss_baseline | CC BY 4.0 (audio clips themselves are CC0) | LICENSE.fuss.txt |
| Russian LibriSpeech / RuLS (speech negatives used for training) | https://www.openslr.org/95 | Public Domain (USA) | LICENSE.ruls.txt |
Notes
- Positive samples (
Привет Аврора) and hard negative phrases were recorded internally for this project and are not distributed with the application. - The trained classifier weights (
models/privet_avrora.onnx,models/privet_avrora_gate.onnx) are provided under the same terms as the application source code (BSD-3-Clause); the licenses above cover the underlying training data and pretrained feature extractor.