Nox Talk: offline voice noise suppression

RNNoise source and default model, Xiph.Org / Jean-Marc Valin / Amazon / Mozilla.
https://github.com/xiph/rnnoise/tree/70f1d256acd4b34a572f999a05c87bf00b67730d
BSD-3-Clause; see licenses/RNNoise-BSD-3-Clause.txt.
Additional per-file BSD notices: licenses/RNNoise-SOURCE-NOTICES.txt.
Portable generic C build; no changes to the model coefficients.

NSNet2 48-kHz baseline model, Microsoft Corporation (DNS Challenge).
https://github.com/microsoft/DNS-Challenge/tree/a052ad5a7714bcc6069b515666e837bf973099de/NSNet2-baseline
Model licensed under Creative Commons Attribution 4.0 International;
see licenses/NSNet2-CC-BY-4.0.txt. Upstream code is MIT-licensed.
Nox modification: the two GRU initial/final states are exposed as graph I/O
for continuous single-frame inference. The trained weights are unchanged.
No endorsement by Microsoft is implied.

DPDFNet2 48-kHz HR, Ceva Inc.
https://huggingface.co/Ceva-IP/DPDFNet
https://github.com/ceva-ip/DPDFNet
Apache License 2.0; see licenses/DPDFNet-Apache-2.0.txt.
Official streaming ONNX export distributed by sherpa-onnx. Weights unchanged.
STFT streaming conventions follow sherpa-onnx (Xiaomi Corporation,
Apache-2.0); Nox uses a Rust FFT implementation and a bounded audio worker.

All models run locally. No microphone audio or model-license telemetry
is sent to the model publishers. Model hashes: manifest.json.
Models are supplied separately from the application source license.
These components are supplied AS IS without warranties; see full licenses.
