# Third-Party Notices

Updated: 2026-10-28 (current model/license inventory; pinned weights unchanged)

## Reviewed production model inventory (2026-10-28)

The current **twelve user-selectable speech-recognition modes** are:
JA Moonshine Small Streaming; DE and ES Moonshine Small Streaming;
ZH, AR, VI and TL Moonshine Tiny Streaming; FR and HI official Vosk Small
0.22; KO SenseVoice Small INT8; TH PengChengStarling INT8; and EN
multilingual Whisper Small. The fixed-language modes do not use Whisper
as a fallback.

Additional model-bearing components include **Ruri v3 70M INT8** for
Japanese semantic classification, **Multilingual E5 Small** for
non-Japanese semantic classification, **Silero VAD** for speech endpoints,
**Kitten TTS Nano 0.8** for optional uncatalogued speech, **CMUDict**
for the optional phonemizer, and pretrained **Qwen3-TTS** used only to
generate the distributed AAC recordings (its model weights are *not*
part of the Web app).

The authoritative current product inventory is
`licenses/CURRENT-MODEL-INVENTORY.json`.
Its `reviewedAt` is our attribution-review date; source model-card
revisions in that file are deliberately pinned, not regenerated to
display a newer date. The reference-only MD5 ledger is
`licenses/CURRENT-LICENSE-DIGESTS.json`.
**MD5 is not used for model integrity; model assets are verified using
pinned SHA-256 hashes.**

**Important qualification — Korean SenseVoice.** The model weights use
the separate FunASR Model Open Source License Agreement (retained v1.1),
not the Sherpa runtime's Apache-2.0 license. The exact converted-weight
permissions and scope are the subject of an unresolved license-owner
clarification (https://github.com/modelscope/FunASR/issues/3715).
We do not claim blanket commercial or redistribution approval.
Thai PengChengStarling's *original* model card labels the checkpoint
Apache-2.0, but the converted repository has no standalone license
metadata; original author, converter and modification provenance are
retained in `PENGCHENG-PROVENANCE.json`. Neither statement certifies
the rights of all training material.

Older historical implementation notes below remain dated history and
must not override this current model list.

This file summarizes third-party software, models, data, and runtime components
used by Mitsukotoba Web. It is intended to make redistribution and attribution
requirements visible. The accompanying `DEPENDENCY_AUDIT.md` records the
GPL/eSpeak-focused technical audit.

## Adopted prerecorded voice

The current fixed English bank uses locally generated Qwen3-TTS AAC audio.
The Web app distributes audio and a provenance manifest; Qwen3 weights and the
MLX runtime are not downloaded to the browser. Kitten remains the optional
uncatalogued free-text synthesis path and the explicit legacy A/B diagnostic.

- Reference: user-selected synthesized VoiceDesign `surprise_02`, seed1202.
- Generation: `mlx-community/Qwen3-TTS-12Hz-1.7B-Base-8bit`, revision
  `e7dd0585652209fa0d7783659aad4e8a324de11c`, seed1202 per utterance.
- Model card license: Apache-2.0; [pinned model card](https://huggingface.co/mlx-community/Qwen3-TTS-12Hz-1.7B-Base-8bit/blob/e7dd0585652209fa0d7783659aad4e8a324de11c/README.md).
- Upstream: https://github.com/QwenLM/Qwen3-TTS
- Full text/audio hashes, reference hash, source commits, and encoding details:
  `assets/voice-1202/manifest.json`.

## GPL / eSpeak NG status

Mitsukotoba Web does not intentionally ship or dynamically import an eSpeak NG
runtime in its production path or development model lab.

The former development-only comparisons using `kokoro-js@1.2.1` and
`@mintplex-labs/piper-tts-web` were removed because those browser paths use
eSpeak NG-based phonemization.

The production Kitten path uses Mitsukotoba's local
`src/vendor/kitten/phonemizer.js`, which is based on CMUDict plus local
conversion/context rules and does not call eSpeak NG. CI rejects known
eSpeak-backed browser runtime identifiers from the deployed source tree.

"GPL-free" here does not mean "permissive-only": the Moonshine WASM provenance
also includes MPL-2.0 and CC BY 4.0 material described below.

## Major dependency summary

| Component | Version / model | License / terms |
| --- | --- | --- |
| Moonshine Voice | 0.1.5 / Japanese and six other fixed-language Streaming | MIT |
| Kitten TTS Nano | 0.8 FP32 | Apache-2.0 |
| CMUDict | pinned 74790861… | BSD-style |
| ONNX Runtime Web | 1.20.0 production Kitten / 1.17.0 model-lab | MIT |
| Eigen provenance | Moonshine / ORT dependency surface | MPL-2.0 |
| VCTK reference clips | Moonshine WASM provenance | CC BY 4.0 |
| Supertonic model-lab comparison | Supertone/supertonic-3 | Supertone OpenRAIL-M |

## Moonshine Voice 0.1.5 / Japanese Streaming STT

Mitsukotoba Web vendors the official Moonshine Voice 0.1.5 WebAssembly runtime
under `src/vendor/moonshine/` and uses Japanese, German, Spanish, Mandarin,
Arabic, Vietnamese and Tagalog Streaming speech-to-text models.

- Upstream: https://github.com/moonshine-ai/moonshine
- Runtime/source default license: MIT
- Streaming STT models: MIT
- Vendored upstream license: `src/vendor/moonshine/LICENSE`

The legacy non-streaming Japanese Base/Tiny models covered by the Moonshine
Community License are not the models used by Mitsukotoba Web.

### Material statically linked by the upstream Moonshine WASM build

Moonshine 0.1.5's upstream WASM build statically links its native core and a
minimal ONNX Runtime build. Relevant audited third-party material includes:

- Eigen 3.4.0 subset — MPL-2.0. Moonshine explicitly builds with
  `EIGEN_MPL2_ONLY=1` and its vendored README states that GPL/LGPL sparse
  modules and wrappers were removed.
- cpp-annote — MIT.
- kaldi-native-fbank — Apache-2.0.
- kissfft — BSD-3-Clause.
- nlohmann/json — MIT.
- ONNX Runtime 1.23.2 — MIT, with its own third-party notices.
- utfcpp — Boost Software License 1.0.
- utf8proc — MIT, with Unicode-derived data subject to the Unicode data notice.
- Silero VAD model/runtime material — MIT.
- Built-in ZipVoice reference clips derived from the VCTK Corpus — CC BY 4.0.

### VCTK attribution

VCTK Corpus, Centre for Speech Technology Research (CSTR), University of
Edinburgh. The VCTK 0.92 corpus is distributed under Creative Commons
Attribution 4.0 International (CC BY 4.0).

- Dataset page: https://datashare.ed.ac.uk/handle/10283/3443
- License: https://creativecommons.org/licenses/by/4.0/

## Kitten TTS Nano 0.8

Mitsukotoba Web retains Kitten TTS Nano 0.8 for uncatalogued free-text synthesis.

- Model: `KittenML/kitten-tts-nano-0.8-fp32`
- Browser integration: adapted browser-only subset of `kitten-tts-js` 0.1.2
- Upstream runtime commit used by the project:
  `222cb5586764fa51f4e73d469d6b1e5d92a56f21`
- License: Apache-2.0
- Upstream: https://github.com/KittenML/KittenTTS

The adapted browser path removes Node-only `fs`, `path`, and `os` code.
Mitsukotoba does not use the eSpeak-backed browser packages removed from the
model lab for this production Kitten path.

## CMU Pronouncing Dictionary (CMUDict)

Mitsukotoba Web uses CMUDict as the primary English pronunciation dictionary
for Kitten TTS.

- Upstream: https://github.com/cmusphinx/cmudict
- Pinned commit: `74790861f652b15e4ac49015a90074ad62a27690`
- License: BSD-style CMUDict license

The dictionary is fetched from the pinned upstream commit and cached by the
browser. Mitsukotoba converts ARPABET entries to the IPA/symbol inventory
expected by Kitten TTS and applies local context/exception rules.

CMUDict copyright notice:

Copyright (C) 1993-2015 Carnegie Mellon University. All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that source redistributions retain the
copyright notice, conditions and disclaimer, and binary redistributions
reproduce them in the documentation and/or other materials.

THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY "AS IS" AND ANY
EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED.

## ONNX Runtime Web used by Kitten and the model lab

The browser Kitten integration loads ONNX Runtime Web 1.20.0. The remaining
Supertonic development comparison currently loads ONNX Runtime Web 1.17.0.

- Upstream: https://github.com/microsoft/onnxruntime
- License: MIT
- Third-party notices: see the matching ONNX Runtime distribution/upstream
  `ThirdPartyNotices.txt`.

## Supertonic development comparison

`model-lab.html` retains a development-only Supertonic 3 comparison.

- Helper source pinned from:
  https://github.com/cskwork/supertonic-tts
- Helper/application code: MIT according to its upstream documentation.
- Model: `Supertone/supertonic-3`
- Model license: Supertone OpenRAIL-M terms apply.
- The pinned helper uses Unicode text preprocessing and does not reference
  eSpeak NG.

This development comparison is not part of the normal Mitsukotoba conversation
path and only loads when the model-lab page is opened and the user runs it.

## Project code

Unless a file or directory says otherwise, Mitsukotoba project code is governed
by the license selected for this repository. Third-party components retain
their own licenses and notices.

## Semantic Lite test edition

The optional Semantic Lite feature runs in a dedicated local browser worker.
It uses the following pinned components:

- Ruri v3 70M, `cl-nagoya/ruri-v3-70m`, revision
  `07a8b0aba47d29d2ca21f89b915c1efe2c23d1cc`: Apache-2.0.
- ModernBERT model code provenance: MIT.
- The source model tokenizer provenance: MIT; this is separate from the JS runtime.
- `@huggingface/tokenizers` 0.2.0 JS runtime: Apache-2.0.
- ONNX Runtime Web 1.30.0: MIT plus its accompanying third-party notices.

The model was exported locally with mean pooling and L2 normalization and
converted to rowwise INT8 token weights plus per-channel INT8 MatMul weights.
Classification uses three frozen logistic regression heads. Project integration
and byte fallback handling were added locally. Generated English is not used.

Complete license copies, the pinned model card and conversion attribution are
in `licenses/semantic/`; model/runtime asset preparation also retains the
applicable copies alongside the prepared assets. ONNX Runtime notices came from
the upstream `v1.30.0` tag:

- https://github.com/microsoft/onnxruntime/blob/v1.30.0/LICENSE
- https://github.com/microsoft/onnxruntime/blob/v1.30.0/ThirdPartyNotices.txt
- https://huggingface.co/cl-nagoya/ruri-v3-70m/tree/07a8b0aba47d29d2ca21f89b915c1efe2c23d1cc

The pinned model and runtime distribution assets are included in the repository
for the opt-in Pages test feature. Local preparation/package scripts preserve
the same hashes and license copies.

## Silero VAD for microphone endpoint detection

- Source: https://github.com/snakers4/silero-vad
- Pinned revision: `1e261b036686cd0017d500ee96acd1c4ba572a9d`.
- License: MIT.
- Bundled ONNX model: 2,327,524 bytes; SHA-256 `1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3`.
- Model and notice are bundled together; see `assets/vad/model-provenance.json` and `LICENSE-SILERO-VAD-MIT.txt` in that directory.
- Runs through the existing local ONNX Runtime Web 1.30.0 distribution used by Semantic. This adds no eSpeak or sherpa TTS dependency.

## Multilingual Whisper Small and E5

The non-Japanese mode uses pinned Xenova Whisper Small (OpenAI MIT;
conversion model card Apache-2.0) and multilingual E5 Small (Microsoft MIT),
Transformers.js 3.8.1 (Apache-2.0), and ONNX Runtime Web 1.30.0 (MIT).
Full copies and pinned attribution are in `licenses/multilingual/NOTICE.md`
and `multilingual-assets/runtime/`. Model weights download only when needed;
all speech recognition and classification run locally. Japanese Moonshine/Ruri
assets and their notices are preserved.

## Thai PengChengStarling

The original model release by PCL-Voice / Yang Bo (stdo) declares Apache-2.0.
The app uses the identifiable INT8 encoder/joiner conversion by Fangjun Kuang
(csukuangfj) / sherpa-onnx, with an ASR-only Apache-2.0 runtime and a local Thai
decoder start-token patch. Complete license, original model card, conversion
sources and modification notices are retained under `licenses/multilingual/`.
See [model notice](licenses/multilingual/PENGCHENG-NOTICE.html) and
[detailed provenance](licenses/multilingual/PENGCHENG-PROVENANCE.json).
This grant concerns the model and the named runtime; it is not permission to
redistribute the original training-code repository, which has no root license.