NOTICE

Vokra
Copyright 2026 ayutaz

This product is licensed under the Apache License, Version 2.0.
See the LICENSE file in the root of this repository.

This NOTICE file records project-wide licensing and distribution decisions.
It is updated whenever a new model, external dependency, or bundled
third-party component with distribution-relevant terms is added.

--------------------------------------------------------------------------
1. BigVGAN (NVIDIA) — MIT attribution
--------------------------------------------------------------------------
CORRECTION (2026-07-22): the NVIDIA BigVGAN reference is **no longer**
non-commercial. github.com/NVIDIA/BigVGAN now ships a standard MIT license
("MIT License / Copyright (c) 2024 NVIDIA CORPORATION"), and the released
weights (e.g. nvidia/bigvgan_v2_24khz_100band_256x) declare `license: mit`.

The clean-room policy below was written when the reference was under the
NVIDIA Source Code License-NC. That premise no longer holds: a direct port
under MIT attribution is now permissible, and the scratch-reimplementation
constraint is lifted. Retained here rather than deleted so the reason the
operator was originally scoped as a from-paper rebuild stays legible.

  BigVGAN
  Copyright (c) 2024 NVIDIA CORPORATION
  Licensed under the MIT License.
  https://github.com/NVIDIA/BigVGAN

The reference-equivalent native operator, strict real-weight binder, and
alias-free activation path landed on 2026-08-21. The complete upstream MIT
text is reproduced in
crates/vokra-ops/THIRD_PARTY_LICENSES/bigvgan-LICENSE.txt.

BigVGAN's `alias_free_activation` implementation is adapted from
`junjun3518/alias-free-torch` under Apache-2.0. Its sinc and low-pass filter
construction additionally derives from Alexandre Defossez's `julius` under
MIT. Vokra's native Rust port changes storage, tensor layout, error handling,
and scalar execution while retaining those notices. The complete license
texts are reproduced in:

  crates/vokra-ops/THIRD_PARTY_LICENSES/alias-free-torch-LICENSE.txt
  crates/vokra-ops/THIRD_PARTY_LICENSES/julius-LICENSE.txt

Sources:
  https://github.com/junjun3518/alias-free-torch
  https://github.com/adefossez/julius

--------------------------------------------------------------------------
2. NVIDIA CUDA / cuDNN / cuBLAS runtime dependencies (not bundled)
--------------------------------------------------------------------------
Vokra does not bundle or redistribute any NVIDIA runtime component
(CUDA runtime / cudart, cuDNN, cuBLAS). When the CUDA backend is enabled,
Vokra detects a system-installed CUDA at runtime via dynamic loading
(e.g. dlopen("libcuda.so") on Linux, LoadLibrary("nvcuda.dll") on Windows).
Installing NVIDIA components and accepting the corresponding NVIDIA EULAs
is the responsibility of the end user or integrating developer. cuDNN is
never a hard requirement of Vokra. The full EULA-compliance strategy (the
5 points: no bundling, runtime detection, explicit version probe, cuDNN
optional, this record) is documented in third_party/NVIDIA-EULA.md, which
is the authoritative reference for the vokra-backend-cuda crate.

Timing note: the CUDA backend is planned for the v0.5-v1.0 releases (M2-03).
The M2-03 foundation slice (crates/vokra-backend-cuda) lands the driver
probe and an FP32 GEMM via raw CUDA Driver API + NVRTC loaded by dlopen,
with NO cudarc/cust/rustacuda binding crate and no bundled CUDA runtime.

--------------------------------------------------------------------------
3. pocketfft (algorithm reference for the CPU FFT implementation)
--------------------------------------------------------------------------
Vokra's CPU FFT lowering (crates/vokra-ops/src/fft/*) is a from-scratch Rust
reimplementation of the algorithm of pocketfft by Martin Reinecke
(https://gitlab.mpcdf.mpg.de/mtr/pocketfft), distributed under the 3-clause
BSD license, Copyright (C) 2010-2019 Max-Planck-Society. The full license
text is reproduced in crates/vokra-ops/THIRD_PARTY_LICENSES/pocketfft-LICENSE.txt.

--------------------------------------------------------------------------
4. piper-plus (MB-iSTFT-VITS2 native TTS and G2P)
--------------------------------------------------------------------------
Vokra's native TTS (crates/vokra-models/src/piper_plus/*) is a from-scratch
Rust reimplementation of the MB-iSTFT-VITS2 inference core of piper-plus
(https://github.com/ayutaz/piper-plus), MIT-licensed. The voice model weights
are converted offline from the upstream piper-plus voices (also MIT). The G2P
(8-language text front-end) is bridged from piper-plus for now; the concrete
reuse form is pending (M0-07-T04, docs/piper-plus-integration.md), so the M0
build ships only a mock phonemizer and does not yet bundle piper-plus G2P.

--------------------------------------------------------------------------
5. Mimi codec (Kyutai / Moshi authors) — CC-BY 4.0 attribution
--------------------------------------------------------------------------
The Mimi RVQ codec integrated by the M3-06 `mimi_rvq` operator
(crates/vokra-ops/src/mimi_rvq.rs) is distributed by Kyutai as part of the
Moshi project under the Creative Commons Attribution 4.0 International
license (CC-BY 4.0). Commercial use is permitted; attribution is required.

  Mimi Codec
  Copyright (c) Kyutai / Moshi authors
  Licensed under CC-BY 4.0
  Source: https://github.com/kyutai-labs/moshi
          https://huggingface.co/kyutai/mimi

The Vokra runtime implements the Mimi decode side (residual codebook
lookup + FP32 residual sum) from published documentation; upstream Mimi
weights are loaded through the standard GGUF path and are subject to the
M2-13 compliance gate (`AttributionRequired` license class).

M4-05 additionally implements the Mimi *neural chain* natively — the
encoder (audio -> RVQ codes: SEANet conv stack + bottleneck transformer +
RVQ quantize) and the neural decoder (RVQ features -> 24 kHz PCM) in
crates/vokra-models/src/mimi/ (shared by Sesame CSM, M4-05, and Moshi,
M4-06). This attribution covers the Kyutai Mimi weights consumed by those
components as well: encoder, quantizer codebooks, and neural decoder
weights all originate from the same CC-BY 4.0 Kyutai release named above.

M4-06 additionally implements the Moshi language model natively — the
Helium temporal transformer, the per-step-weight depformer, the
inner-monologue text stream, and the full-duplex session runtime in
crates/vokra-models/src/moshi/ (whisper.cpp-style re-implementation from
the published kyutai-labs/moshi sources; no upstream code is copied).
The Moshi LM weights (kyutai/moshiko-pytorch-bf16 `model.safetensors`)
are likewise Kyutai CC-BY 4.0 — this attribution covers them, and the
runtime additionally surfaces a deployer-facing attribution string
(GGUF `vokra.provenance.attribution` -> `Session::attribution` /
`vokra_model_attribution` / the vokra-cli banner) so downstream
applications can meet the display obligation (FR-MD-09).

The Kyutai STT incremental language-model component in
`crates/vokra-models/src/kyutai_stt/streaming_lm.rs` is an original
first-party Rust reimplementation of the published Moshi `LMGen` streaming
semantics at commit `e6a55d2722a65870ef52a6c9f6ecfc0e90f38362`; no literal
upstream source code is copied.  The pinned Moshi source package is
MIT-licensed (pinned `moshi/LICENSE`).  This attribution records the source of the
implementation's semantics; it does not claim that a license attaches to
mathematical semantics.  This component does not by itself grant
redistribution rights for Kyutai STT weights, Mimi assets, or a complete
speech pipeline.

This attribution is separate from the EU AI Act §50 AI-generated-audio
disclosure obligation. The disclosure obligation is enforced at output
time by the runtime's `WatermarkConfig` API and is orthogonal to this
codec-level credit.

--------------------------------------------------------------------------
6. EnCodec (Meta) — pretrained weights permanently excluded (FR-OP-32)
--------------------------------------------------------------------------
Pretrained EnCodec weights (Meta / Facebook Research) are licensed under
CC-BY-NC 4.0 (non-commercial). Per FR-OP-32 (permanent constraint) and
docs/license-audit.md §3, no EnCodec weight file ships with Vokra's
official model zoo. The `mimi_rvq` op is codebook-shaped and could host
an EnCodec code path, but any EnCodec use requires the caller to bring
their own weight file and opt in through the M2-13 compliance research
flag (`VOKRA_ALLOW_RESEARCH_LICENSE=1` /
`CompliancePolicy::with_research_license(true)` /
`ComplianceLevel::Research`) — the runtime never loads them by default.
Release CI enforces the distribution-side exclusion via
`scripts/compliance/check-encodec-exclusion.sh`. Commercial codec
alternatives supported by Vokra: DAC (MIT), Mimi (CC-BY 4.0), WavTokenizer
(MIT), X-Codec 2 (MIT).

--------------------------------------------------------------------------
7. SpeexDSP echo canceller (Rust port of mdf.c for the `aec` operator)
--------------------------------------------------------------------------
Vokra's acoustic echo cancellation operator (crates/vokra-ops/src/aec.rs,
M4-03 / FR-OP-60) is a from-scratch Rust port — a derivative work, not a
clean-room reimplementation — of the SpeexDSP echo canceller
(libspeexdsp/mdf.c, the MDF/AUMDF adaptive filter with the Valin 2007
variable learning rate), Copyright (C) 2003-2008 Jean-Marc Valin,
distributed under the 3-clause BSD license as part of SpeexDSP
(Copyright 2002-2008 Xiph.org Foundation and contributors).

  Source: https://github.com/xiph/speexdsp
          (pinned commit 7a158783df74efe7c2d1c6ee8363c1e695c71226)
  License: 3-clause BSD — full text reproduced in
          crates/vokra-ops/THIRD_PARTY_LICENSES/speexdsp-LICENSE.txt

No SpeexDSP C source is bundled, compiled, or linked into any Vokra
artifact (NFR-DS-02 zero-dependency); the port is pure Rust. The upstream
code is only executed offline by the developer-side parity tool
(tools/parity/aec_dump.py) to produce the committed reference fixtures
under tests/parity/aec/. The far-end reference queue
(crates/vokra-core/src/stream/aec_ref.rs) is an original Vokra design and
not derived from SpeexDSP.

--------------------------------------------------------------------------
8. DeepFilterNet (Rikorose) — MIT attribution (M4-20 denoise op)
--------------------------------------------------------------------------
The M4-20 `denoise` operator (crates/vokra-ops/src/denoise.rs, FR-OP-61)
reimplements the DeepFilterNet speech-enhancement topology (ERB gain + deep
filtering) from scratch in Rust based on the published architecture:

  DeepFilterNet
  Copyright (c) Hendrik Schröter (Rikorose) and contributors
  Licensed under the MIT License.
  https://github.com/Rikorose/DeepFilterNet

MIT permits commercial use; attribution (this notice) is retained. Vokra
loads only converted weights (offline, to the `vokra.denoise.*` GGUF); no
ONNX / PyTorch runtime dependency is introduced. The MIT license text is to
be reproduced in crates/vokra-ops/THIRD_PARTY_LICENSES/deepfilternet-LICENSE.txt
when the real DeepFilterNet checkpoint is added to the model zoo (owner
sign-off, docs/license-audit.md §3 — M4-20 T18). GTCRN (MIT) is an alternate
denoise candidate; its attribution is added here if/when its native forward is
integrated.

--------------------------------------------------------------------------
8a. RNNoise v0.2 (Xiph.Org) — BSD-3-Clause attribution
--------------------------------------------------------------------------
Vokra's `rnnoise` model runtime is a pure-Rust implementation of the RNNoise
v0.2 signal-processing and recurrent-network topology. The official source is
also compiled only by developer-side parity oracles; no RNNoise C source or
library is bundled, compiled, or linked into a Vokra runtime artifact.

  RNNoise v0.2
  Copyright (c) 2007-2017, 2024 Jean-Marc Valin
  Copyright (c) 2023 Amazon
  Copyright (c) 2017, Mozilla
  Copyright (c) 2005-2017, Xiph.Org Foundation
  Copyright (c) 2003-2004, Mark Borgerding
  Licensed under the BSD-3-Clause License.
  https://github.com/xiph/rnnoise/releases/tag/v0.2

The complete upstream license is reproduced in
crates/vokra-models/THIRD_PARTY_LICENSES/rnnoise-LICENSE.txt.

--------------------------------------------------------------------------
10. Japanese G2P chain (Open JTalk lineage) — BSD-3-Clause notices
--------------------------------------------------------------------------
Vokra's Japanese text-to-speech front-end resolves grapheme-to-phoneme
through `jpreprocess`, a Rust reimplementation of Open JTalk that bundles
the NAIST Japanese Dictionary. It is used by the excluded workspace
`integrations/vokra-piper-g2p`, which drives plain-Japanese synthesis for
the piper-plus voices.

That lineage is permissively licensed end to end -- there is no GPL
obligation anywhere in it, because the HTS Working Group elected the BSD
arm of MeCab's disjunctive tri-license upstream. Vokra makes the same
election explicitly, and records it here so an auditor does not have to
infer which branch was taken:

  **Vokra elects the BSD-3-Clause arm of the MeCab tri-license
  (GPL-2.0 / LGPL-2.1 / BSD-3-Clause).**

BSD-3-Clause clause 2 requires that a *binary* redistribution reproduce the
copyright notice and conditions. Vokra ships binaries, so these notices are
carried here rather than left to the upstream packaging -- which does not
discharge the obligation on our behalf.

  jpreprocess
  Licensed under the BSD-3-Clause License.
  https://github.com/jpreprocess/jpreprocess

  Open JTalk
  Copyright (c) 2008-2016 Nagoya Institute of Technology
                          Department of Computer Science
  Licensed under a Modified BSD (3-clause) License.
  http://open-jtalk.sourceforge.net/

  hts_engine API
  Copyright (c) 2001-2015 Nagoya Institute of Technology
                          Department of Computer Science
  Copyright (c) 2001-2008 Tokyo Institute of Technology
                          Interdisciplinary Graduate School of
                          Science and Engineering
  Licensed under a Modified BSD (3-clause) License.
  http://hts-engine.sourceforge.net/

  MeCab
  Copyright (c) 2001-2008, Taku Kudo
  Copyright (c) 2004-2008, Nippon Telegraph and Telephone Corporation
  Tri-licensed GPL-2.0 / LGPL-2.1 / BSD-3-Clause; Vokra uses it under
  BSD-3-Clause (see the election above).
  https://taku910.github.io/mecab/

  NAIST Japanese Dictionary (NAIST-JDic)
  Copyright (c) 2009, Nara Institute of Science and Technology
  Licensed under a Modified BSD (3-clause) License.

  UniDic (as vendored in the Open JTalk lineage)
  Copyright (c) The UniDic Consortium / NINJAL
  The artifacts vendored in this chain are BSD-3-Clause.
  NOTE: NINJAL's own direct UniDic distribution is tri-licensed
  GPL-2.0 / LGPL-2.1 / Modified BSD. Anything that pulls UniDic directly
  (rather than through this chain) must make the BSD election explicitly.
  https://clrd.ninjal.ac.jp/unidic/

Deliberately NOT used, and why:

  - **eSpeak-NG (GPL-3.0)** is never linked. The common Python route reaches
    it through `phonemizer` via in-process `ctypes` dynamic linking, which is
    the strongest form of the copyleft argument -- not a subprocess shell-out.
  - **num2words (LGPL-2.1+)** is not used in any Japanese numeral path. It is
    trivially reimplementable and its PyPI metadata declares a bare "LGPL"
    with no version, which also defeats automated scanners.
  - **pyopenjtalk's bundled `mei_normal.htsvoice`** (CC BY 3.0, MMDAgent
    Project Team) is not redistributed. Its attribution requirement attaches
    to synthesis and wheel redistribution, neither of which Vokra performs.

The BSD-3-Clause license texts are to be reproduced under
integrations/vokra-piper-g2p/THIRD_PARTY_LICENSES/ when the Japanese G2P
path is included in a binary release artifact.

--------------------------------------------------------------------------
11. NVIDIA TitaNet-Large (speaker verification) — CC-BY-4.0 attribution
--------------------------------------------------------------------------
The `ModelKind::TitaNet` converter (crates/vokra-convert/src/models/titanet.rs,
2026-07-30) writes GGUFs from the upstream NVIDIA TitaNet-Large speaker-
verification checkpoint. Model weights are distributed by NVIDIA under
the Creative Commons Attribution 4.0 International license (CC-BY 4.0).
Commercial use is permitted; attribution is required.

  TitaNet-Large (speaker verification)
  Copyright (c) NVIDIA CORPORATION
  Licensed under CC-BY 4.0
  Source: https://huggingface.co/nvidia/speakerverification_en_titanet_large

The converter additionally stamps a runtime-side FR-MD-09 attribution
string (`vokra.provenance.attribution` -> Session::attribution /
`vokra_model_attribution` / the vokra-cli banner) so downstream
applications can meet the CC-BY 4.0 display obligation without having
to know NVIDIA's terms independently.

Runtime port (native forward pass for `TITANET_SPEAKER_ENCODE_OP`,
FR-OP-80 variant) is **out-of-scope** for the converter landing: the
op-kind identifier is reserved in crates/vokra-core/src/m5_residual_ops.rs
as M5-residual, and CAM++ (Apache-2.0, no attribution overhead) already
covers the fbank-80 -> 192-d speaker-embedding functional surface via
`vokra-models::speaker_encode`. NVIDIA credit through this NOTICE +
the FR-MD-09 attribution chunk applies whenever a TitaNet GGUF is
redistributed as a Vokra artifact; a downstream that never redistributes
the TitaNet weights does not incur the obligation.

CC-BY 4.0 attribution is separate from the EU AI Act §50 AI-generated-
audio disclosure obligation (TitaNet is a speaker-verification model,
not an audio generator, so §50 does not apply here).

The CC-BY 4.0 license text is to be reproduced in
crates/vokra-convert/THIRD_PARTY_LICENSES/titanet-CC-BY-4.0.txt when
the TitaNet checkpoint is added to the official model zoo (owner
sign-off, docs/license-audit.md §3.1 = 2026-07-30 yousan).

--------------------------------------------------------------------------
12. Pipecat SmartTurn v2 — BSD-2-Clause notice
--------------------------------------------------------------------------
Vokra's SmartTurn v2 converter, native endpoint classifier, and independent
reference sidecar target Pipecat's released `pipecat-ai/smart-turn-v2`
checkpoint and the endpoint head published in the Pipecat repository at
commit c560a748b4213ca8db6f43a5d165d91aaa124a52.

  Pipecat / SmartTurn
  Copyright (c) 2024–2026, Daily
  Licensed under the BSD 2-Clause License.
  Source: https://github.com/pipecat-ai/pipecat
          https://huggingface.co/pipecat-ai/smart-turn-v2

The runtime remains dependency-free and does not compile or link Pipecat,
Transformers, ONNX Runtime, or Python. The pinned Python sidecar is an offline
parity oracle only. Official model weights are not bundled or published by
this change. The complete license text is reproduced in
tools/parity/THIRD_PARTY_LICENSES/pipecat-LICENSE.txt.

--------------------------------------------------------------------------
13. TEN-VAD v1.0 — restricted upstream license and LPCNet notices
--------------------------------------------------------------------------
Vokra's strict converter, native streaming VAD, and independent parity
sidecar target TEN-framework/ten-vad commit
8e96899ba05a8e8c0e883ec7417e7a144bd9dec0 (tag v1.0-ONNX).

The upstream LICENSE is not unmodified Apache-2.0. It adds non-compete,
application-only deployment conditions and requires derivative works to remain
under those terms. Canonical GGUF conversion therefore stamps
`LicenseRef-Agora-TEN-VAD-Open-Source-License-2025` as
`redistribution-forbidden`; this repository does not bundle or publish the
official weights. Application-local use remains subject to the upstream terms.

The LPCNet-derived frontend is separately covered by upstream NOTICES:

  Copyright (c) 2017-2019 Mozilla                     (BSD-2-Clause)
  Copyright (c) 2017-2018 Mozilla                     (BSD-3-Clause)
  Copyright (c) 2007-2017 Jean-Marc Valin             (BSD-3-Clause)
  Copyright (c) 2005-2017 Xiph.Org Foundation         (BSD-3-Clause)
  Copyright (c) 2003-2004 Mark Borgerding             (BSD-3-Clause)

The complete upstream license and both BSD notices are reproduced in
tools/parity/THIRD_PARTY_LICENSES/ten-vad-LICENSE.txt and
tools/parity/THIRD_PARTY_LICENSES/ten-vad-NOTICES.txt.

--------------------------------------------------------------------------
14. FunASR FSMN-VAD — MIT code notice
--------------------------------------------------------------------------
Vokra's native FSMN-VAD encoder and online frontend follow the released
FunASR implementation pinned at commit
3c58cb56a56598232c3efffa15d313d7e82a4307.

  FunASR
  Copyright (c) 2025 FunASR
  Licensed under the MIT License.
  Source: https://github.com/modelscope/FunASR

The complete MIT text is reproduced in
crates/vokra-ops/THIRD_PARTY_LICENSES/funasr-LICENSE.txt. The official
`funasr/fsmn-vad` checkpoint is separately licensed Apache-2.0; code and
weight provenance are not conflated. No official weight is bundled in this
repository.

--------------------------------------------------------------------------
15. NVIDIA Parakeet-CTC-1.1B — CC-BY-4.0 attribution
--------------------------------------------------------------------------
Vokra's strict converter and native ASR runtime target the authenticated
`nvidia/parakeet-ctc-1.1b` release. The model weights are distributed by
NVIDIA under the Creative Commons Attribution 4.0 International license.
Commercial use is permitted; attribution is required.

  NVIDIA Parakeet-CTC-1.1B
  Copyright (c) NVIDIA CORPORATION
  Licensed under CC-BY 4.0
  Source: https://huggingface.co/nvidia/parakeet-ctc-1.1b

The converter embeds this notice in `vokra.provenance.attribution` and stamps
the canonical `attribution-required` weight class. The runtime is a native,
zero-dependency Rust implementation and does not bundle or link NVIDIA CUDA,
cuDNN, PyTorch, Transformers, ONNX Runtime, Python, or upstream source code.
Transformers is used only as the pinned offline reference oracle that produced
the committed numerical fixtures. This runtime change did not upload or
replace any model artifact.

--------------------------------------------------------------------------
16. NVIDIA NanoCodec — NVIDIA Open Model License attribution
--------------------------------------------------------------------------
The NVIDIA NeMo NanoCodec 22.05 kHz checkpoint family is distributed under
the NVIDIA Open Model License Agreement, Version Release Date June 14, 2024.
Commercial use and redistribution are permitted subject to that agreement.
Section 3.1 requires every redistribution to include the agreement and the
following exact sentence in a Notice text file:

Licensed by NVIDIA Corporation under the NVIDIA Open Model License

Source model family:
  https://huggingface.co/nvidia/nemo-nano-codec-22khz-0.6kbps-12.5fps

The controlling June-2024 agreement is vendored at
third_party/nvidia-open-model-license-agreement-june-2024.txt. Its official
source PDF SHA-256 is
396d7de220f6b0e6fcfe33836f1b99a9943769b97c56330444f8efe1b441f607.
The current NVIDIA web agreement was revised on October 24, 2025; NanoCodec's
official model card links the June-2024 PDF, so distributors must not silently
substitute the later text without re-auditing the checkpoint's governing
terms.

--------------------------------------------------------------------------
17. Meta Audiobox Aesthetics — CC-BY-4.0 attribution
--------------------------------------------------------------------------

Vokra's strict converter and native CPU/Metal scorer target the immutable
`facebook/audiobox-aesthetics` release. The model weights are distributed
under the Creative Commons Attribution 4.0 International license. Commercial
use is permitted; attribution is required.

  Meta Audiobox Aesthetics
  Copyright (c) Meta Platforms, Inc. and affiliates
  Licensed under CC-BY 4.0
  Source: https://github.com/facebookresearch/audiobox-aesthetics
          https://huggingface.co/facebook/audiobox-aesthetics

The native Rust runtime reimplements the WavLM backbone and the four learned
CE / CU / PC / PQ quality heads without linking PyTorch, ONNX Runtime or any
upstream runtime dependency. The converter embeds the attribution text and
canonical `attribution-required` weight class into the GGUF. No upstream
weight is bundled in this source repository.

--------------------------------------------------------------------------
17a. Meta AudioSeal — MIT
--------------------------------------------------------------------------

Vokra's strict converter and native CPU/Metal watermark runtime target the
four official `facebook/audioseal` generator/detector checkpoints at immutable
Hugging Face revision 3c19eba53390776cf2cc9ed5f6c9ac67ce72ecba and source
revision e63a8a0e5cdf7bb797159c92ba15961557fe9bd2. The source code and model
weights are distributed under the MIT License.

  Meta AudioSeal
  Copyright (c) Meta Platforms, Inc. and affiliates
  Licensed under the MIT License
  Source: https://github.com/facebookresearch/audioseal
          https://huggingface.co/facebook/audioseal

The native Rust runtime independently implements the complete-buffer base and
streaming-trained SEANet generator/detector forwards without linking PyTorch,
ONNX Runtime or another upstream runtime dependency. No upstream weight is
bundled in this source repository.

--------------------------------------------------------------------------
17b. Microsoft DNS-Challenge NSNet2 — CC-BY-4.0 attribution
--------------------------------------------------------------------------

Vokra's NSNet2 converter and native CPU/Metal runtime target the 20 ms
baseline model at Microsoft DNS-Challenge commit
8b87a33b2892f147b5c7ad39ea978453730db269. That revision licenses source code
under MIT in `LICENSE-CODE`, while the root `LICENSE` and README Legal Notices
place documentation and other released content under the Creative Commons
Attribution 4.0 International license. Vokra therefore treats the released
ONNX model and converted weights as CC-BY-4.0 content. Commercial use is
permitted; attribution is required.

  NSNet2 20 ms baseline model content
  Creator: Microsoft DNS-Challenge project
  Licensed under CC-BY 4.0
  Source: https://github.com/microsoft/DNS-Challenge/tree/8b87a33b2892f147b5c7ad39ea978453730db269/NSNet2-baseline
  License: https://github.com/microsoft/DNS-Challenge/blob/8b87a33b2892f147b5c7ad39ea978453730db269/LICENSE

Changes by Vokra: the ONNX initializer payload is converted to GGUF; numeric
initializer names and matrix axes may be normalized to the native runtime
schema. The native Rust signal-processing and inference implementation does
not link or bundle Microsoft code, ONNX Runtime, or protobuf. No upstream
weight is bundled in this source repository.

--------------------------------------------------------------------------
17c. Meta Facebook Denoiser DNS48 — CC-BY-NC-4.0 research-only
--------------------------------------------------------------------------

Vokra's strict converter and native CPU/Metal speech-enhancement runtime
target the causal DNS48 release from `facebookresearch/denoiser` at source
revision 8afd7c166699bb3c8b2d95b6dd706f71e1075df0. The source and released
weights are distributed under the Creative Commons Attribution-
NonCommercial 4.0 International license. Commercial use is not authorized.

  Meta Facebook Denoiser DNS48
  Copyright (c) Facebook, Inc. and its affiliates
  Licensed under CC-BY-NC 4.0
  Source: https://github.com/facebookresearch/denoiser
  Paper: https://arxiv.org/abs/2006.12847

The native Rust runtime reimplements the official waveform U-Net, causal
LSTM and fixed sinc resampling forward without linking PyTorch, ONNX Runtime
or upstream code. No upstream checkpoint is bundled in this source
repository. Converted weights remain `non-commercial`; default loading and
Hugging Face publication require the explicit research/non-commercial opt-in
documented by Vokra's compliance gate.

--------------------------------------------------------------------------
17d. NVIDIA Canary-1B-Flash — CC-BY-4.0 attribution
--------------------------------------------------------------------------

Vokra's strict converter and native CPU/Metal ASR/AST runtime target the
immutable `nvidia/canary-1b-flash` release at Hugging Face revision
2b6e4d2dacb11cc1b1724de31bb48fe68c26c12e. NVIDIA distributes the model
weights under the Creative Commons Attribution 4.0 International license.
Commercial use is permitted; attribution is required.

  NVIDIA Canary-1B-Flash
  Copyright (c) NVIDIA CORPORATION
  Licensed under CC-BY 4.0
  Source: https://huggingface.co/nvidia/canary-1b-flash

The native Rust runtime independently implements the released FastConformer
encoder, Transformer AED decoder, Canary2 prompt and aggregate tokenizer. It
does not link PyTorch, NVIDIA NeMo, ONNX Runtime or another upstream runtime
dependency. The converter stamps this attribution and the
`attribution-required` weight class into each complete GGUF. No upstream
checkpoint is bundled in this source repository, and this source change does
not upload or replace a model artifact.

--------------------------------------------------------------------------
17e. FireRedASR AED beam semantics — Apache-2.0 source reference
--------------------------------------------------------------------------

Vokra's first-party Rust FireRed AED implementation uses the published
TransformerDecoder beam-search semantics from FireRedTeam/FireRedASR at
source revision 834635e4cf277ed8ca92049fc375b17c3dc20748. The fixed source
repository's LICENSE is Apache-2.0; its complete source tree has no NOTICE
file. This attribution identifies the source of the algorithmic contract,
not a bundled Python runtime or model checkpoint.

  FireRedTeam/FireRedASR
  Source: https://github.com/FireRedTeam/FireRedASR/tree/834635e4cf277ed8ca92049fc375b17c3dc20748
  License: https://github.com/FireRedTeam/FireRedASR/blob/834635e4cf277ed8ca92049fc375b17c3dc20748/LICENSE

Changes by Vokra: independent Rust storage, explicit error handling,
Compute-dispatched operators, incremental projected-key/value caches and
synthetic mechanism tests. Source inspection and this attribution do not
approve dependency payloads, training provenance, checkpoint execution,
redistribution or publication. Real-weight and Apple parity remain separate
gates.

--------------------------------------------------------------------------
18. Maintaining this file
--------------------------------------------------------------------------

Any pull request that adds support for a new model or introduces a new
external dependency must run the license audit (docs/license-audit.md) and,
when the addition carries attribution or distribution-relevant terms,
append the corresponding notice to this file in the same pull request.
See CONTRIBUTING.md for the full policy.