ThoxSpeech
Copyright (c) 2026 THOX.ai LLC.

This product is a rebranded derivative work. It bundles, links against, or
derives from the third-party components listed below. All upstream copyright
notices and license texts are preserved in this repository and MUST be retained
in any redistribution, in source or binary form.

================================================================================
1. VibeASR.cpp  --  PRIMARY UPSTREAM (this repository is a fork of it)
================================================================================
Upstream:   https://github.com/microsoft/VibeASR.cpp
Copyright:  (c) Microsoft Corporation.
License:    MIT  --  full text in LICENSE, verbatim copy in THIRD_PARTY_NOTICES.md

ThoxSpeech is a fork of VibeASR.cpp. The heterogeneous-quantization inference
core -- the I8_S VAE tokenizer path, the I2_S BitNet LM path, the custom SIMD
multiply-add-dot kernels, and the ggml operator fusion -- is Microsoft's work,
used here unmodified in substance.

Significant changes made by THOX.ai LLC are enumerated in CHANGES.md.

================================================================================
2. VibeVoice-ASR-BitNet  --  MODEL WEIGHTS
================================================================================
Upstream:   https://huggingface.co/microsoft/VibeVoice-ASR-BitNet
Copyright:  (c) Microsoft Corporation.
License:    MIT (not gated)

The quantized GGUF artifacts distributed as "ThoxSpeech ASR" weights are
byte-identical mirrors of the upstream VibeVoice-ASR-BitNet GGUF files. THOX.ai
did not retrain, fine-tune, or re-quantize these weights. See MODEL_CARD.md.

Cite the original authors:

  @article{xu2025vibeasrbitnet,
      title={VibeVoice-ASR-BitNet Technical Report},
      author={Xu, Songchen and Song, Ting and Huang, Shaohan and Peng, Zhiliang
              and Xia, Yan and Tu, Yujie and Huang, Xin and Yu, Jianwei and
              Dong, Li and Wei, Furu},
      journal={arXiv preprint arXiv:2607.21075},
      year={2025}
  }

================================================================================
3. Qwen2.5-1.5B  --  BASE LANGUAGE MODEL OF THE ASR DECODER
================================================================================
Upstream:   https://huggingface.co/Qwen/Qwen2.5-1.5B
Copyright:  (c) Alibaba Cloud / Qwen Team.
License:    Apache License 2.0  --  verbatim copy in THIRD_PARTY_NOTICES.md

The ASR decoder is derived from Qwen2.5-1.5B, which replaces the Qwen2.5-7B
decoder used by the full-size VibeVoice-ASR model. Apache-2.0 section 4(b)
requires derivative works to carry prominent notice of modification: the model
has been quantized to I2_S (ternary BitNet) weights with Q6_K token embeddings
and converted to GGUF. Apache-2.0 sections 4(c) and 4(d) require retention of
this NOTICE file and the license text in redistributions.

================================================================================
4. llama.cpp / ggml  --  INFERENCE FRAMEWORK (git submodule, 3rdparty/llama.cpp)
================================================================================
Upstream:   https://github.com/XsquirrelC/llama.cpp (branch vibeasr-dev)
Ancestry:   ggml-org/llama.cpp -> Eddie-Wang1120/llama.cpp -> XsquirrelC/llama.cpp
Copyright:  (c) 2023-2026 The ggml authors.
License:    MIT

Vendored as a git submodule at the pinned vibeasr-dev branch. Its own LICENSE
file travels with it and is not modified by this repository.

================================================================================
5. dr_wav.h  --  WAV DECODER (utils/dr_wav.h)
================================================================================
Upstream:   https://github.com/mackron/dr_libs
Copyright:  (c) David Reid.
License:    public domain (Unlicense) OR MIT-0, at your option -- header retains
            its own embedded license block, unmodified.

================================================================================
TRADEMARK NOTICE
================================================================================
"ThoxSpeech", "ThoxRoute", "ThoxOS" and "THOX" are marks of THOX.ai LLC.

Neither the MIT License nor the Apache License 2.0 grants trademark rights.
This repository does NOT use the "Microsoft", "VibeVoice", "VibeASR", "Qwen",
or "Alibaba" names to endorse or promote ThoxSpeech, and THOX.ai LLC claims no
affiliation with, sponsorship by, or endorsement from Microsoft Corporation or
Alibaba Cloud. Upstream names appear here solely as factual attribution of
provenance, which is what the licenses require.