OFFLINE AI NPC / DOCS
PDF DEMO BUY — $89

Third-party notices — Offline AI NPC

Every third-party component the plugin touches, whether it is bundled with the package, required as a dependency the customer installs, or downloaded at runtime. Missing third-party licences is a documented Asset Store rejection reason, so this file is maintained from stage 0 and updated by every stage that adds a dependency.

Status legend for "Verified" yes — licence text read from the shipped copy or the package manifest in this repo. TODO — the licence is publicly stated but has not yet been read and archived here. Everything marked TODO must be resolved before submission.

Runtime dependencies the customer installs (never bundled)

Bundling either of these causes duplicate-assembly compile errors for anyone who already has them, so the plugin declares them as dependencies and guards every integration point with a version define (see Core/DependencyGuard.cs).

ComponentAuthorLicenceSourceHow it is usedVerified
LLM for Unity (LLMUnity)UndreamAIApache-2.0https://github.com/undreamai/LLMUnityDependency. C# bindings the host uses to run the local LLM. Guarded by LLMUNITY_PRESENT.yes (Assets/LLMUnity/LICENSE.md, v3.0.3)
LlamaLibUndreamAIMIThttps://github.com/undreamai/LlamaLibNative binaries shipped inside LLMUnity (llama.cpp server build). Not redistributed by this plugin.TODO
llama.cppGeorgi Gerganov and contributorsMIThttps://github.com/ggerganov/llama.cppInference engine underneath LlamaLib.TODO
whisper.unityMacoronMIThttps://github.com/Macoron/whisper.unityDependency. Unity bindings for speech-to-text. Guarded by WHISPER_PRESENT.yes (package.json, v1.4.0)
whisper.cppGeorgi Gerganov and contributorsMIThttps://github.com/ggerganov/whisper.cppNative STT engine underneath whisper.unity.TODO
Newtonsoft Json for UnityUnity Technologies / James Newton-KingMITcom.unity.nuget.newtonsoft-jsonDependency (UPM). Serialises conversation memory.TODO

External processes the plugin drives (not redistributed)

These are separate executables the user installs. The plugin starts, talks to, and shuts them down; it does not ship their binaries.

The espeak-ng row below is the one that decides whether this package can ship at all. What "external process" has to mean in practice, what a buyer owes if they redistribute Piper, and what this demo project itself owes, are all in ESPEAK_BOUNDARY.md. PackagingBoundaryTests fails the build if any of it lands under Assets/OfflineAINPC/.

ComponentAuthorLicenceSourceHow it is usedVerified
PiperMichael Hansen (rhasspy)MIThttps://github.com/rhasspy/piperExternal process. One piper.exe per voice model, driven over stdin JSON.TODO
piper-phonemizeMichael Hansen (rhasspy)MIThttps://github.com/rhasspy/piper-phonemizeShips with Piper; converts text to phonemes.TODO
espeak-ngReece Dunn and contributorsGPL-3.0https://github.com/espeak-ng/espeak-ngUsed by piper-phonemize inside the Piper process. Copyleft — treat as a hard constraint: it must stay an external process the user installs, never a bundled or linked component of the package. The package ships none of it; the interview demo build does, and discharges GPL-3.0 as set out below.yes (GPL-3.0.txt, version 1.52.0)
ONNX RuntimeMicrosoftMIThttps://github.com/microsoft/onnxruntimeShips with Piper; runs the voice models.TODO
VOICEVOX ENGINEHiroshiba Kazuyuki and contributorsLGPL-3.0 (engine)https://github.com/VOICEVOX/voicevox_engineExternal HTTP server the plugin queries for Japanese TTS. User-installed and user-managed.TODO
VOICEVOX character voicesper-character rights holdersper-character termshttps://voicevox.hiroshiba.jp/term/Generated audio only; each character has its own credit/usage terms the END USER accepts. The plugin must surface, not absorb, those terms.TODO

GPL-3.0 compliance for the interview demo build

The package ships no speech engine. The interview demo, distributed as a player, does: piper.exe, piper_phonemize.dll and espeak-ng.dll all sit in its StreamingAssets. espeak-ng is GPL-3.0, so that build — not the package — carries the obligations.

What GPL-3.0 asksWhere it is met
A copy of the licence, naming what it coversGPL-3.0.txt, staged into the build's StreamingAssets/LICENSES beside this file
Corresponding source for the exact binary shippedBelow. A release tag, not "latest"
No further restrictions on that componentThe demo has no EULA; nothing forbids replacing or redistributing espeak-ng

What is actually in the build, and where its source is:

BinaryVersionSource
piper.exe, piper_phonemize.dllPiper 1.2.0 (release 2023.11.14-2)https://github.com/rhasspy/piper/releases/tag/2023.11.14-2
espeak-ng.dll, espeak-ng-data/1.52.0, from the fork piper-phonemize builds againsthttps://github.com/rhasspy/espeak-ng — commit 0f65aa301e0d6bae5e172cc74197d32a6182200f

The binaries are redistributed unmodified: they are the upstream release artefacts, copied in, never rebuilt or patched here. Anyone who receives the build may request the corresponding source and will be pointed at the URLs above; if either upstream disappears, an archive of that exact commit has to be hosted alongside the download instead.

Version strings were read out of the shipped binaries rather than taken from a changelog, because the changelog describes what upstream released and this table has to describe what is in this build.

Model weights (downloaded or supplied by the user, not shipped)

Dated copies of each licence and model card the demo depends on are in model-cards-2026-08-16/, taken from the repositories the weights are actually pulled from. A quantisation repository can carry terms of its own on top of the base model, so the copy is of the repository that serves the file rather than of the model it was made from — here both are Qwen's own, both Apache-2.0.

ComponentAuthorLicenceSourceHow it is usedVerified
Qwen3-1.7B (GGUF, Q8_0)Alibaba Cloud / Qwen teamApache-2.0https://huggingface.co/Qwen/Qwen3-1.7B-GGUFThe interview demo's default. 1 834 426 016 bytes, fetched on first run into the player's data folder.yes (model-cards-2026-08-16/)
Qwen3-4B (GGUF, Q4_K_M)Alibaba Cloud / Qwen teamApache-2.0https://huggingface.co/Qwen/Qwen3-4B-GGUFThe demo's opt-in upgrade, 2 497 280 256 bytes. Also what Yume no Ma itself runs.yes (model-cards-2026-08-16/)
Qwen2.5-7B-Instruct (GGUF, Q4_K_M)Alibaba Cloud / Qwen teamApache-2.0https://huggingface.co/Qwen/Qwen2.5-7B-InstructDialogue model for projects that want it. User-supplied.TODO
Whisper ggml models (tiny / tiny.en / base.en / small)OpenAI, ggml conversions by Georgi GerganovMIThttps://huggingface.co/ggerganov/whisper.cppSpeech-to-text weights. User-supplied.TODO
Piper voice en_US-amy-mediumPiper / rhasspy voice collectionper-datasethttps://huggingface.co/rhasspy/piper-voicesEnglish voice model. User-supplied. Per-voice dataset licence must be listed individually.TODO
Piper voice en_US-hfc_female-mediumPiper / rhasspy voice collectionper-datasethttps://huggingface.co/rhasspy/piper-voicesSecond English voice.TODO
Piper voice es_ES-davefx-mediumPiper / rhasspy voice collectionper-datasethttps://huggingface.co/rhasspy/piper-voicesSpanish voice.TODO
Piper voice ru_RU-dmitri-mediumPiper / rhasspy voice collectiondataset CC0-1.0, provenance disputedhttps://huggingface.co/rhasspy/piper-voicesRussian voice, used by Yume no Ma. Stage 5E recorded it as cleared for commercial redistribution on the dataset line alone; VOICE_LICENCES.md withdrew that on the Training line — the card says it is finetuned from lessac, whose dataset forbids commercial use. Not shipped by the plugin or the demo.see VOICE_LICENCES.md
Piper voice zh_CN-chaowen-mediumPiper / rhasspy voice collectionper-datasethttps://huggingface.co/rhasspy/piper-voicesChinese voice, present on this machine for the stage-5F phonemiser experiment. Not a configured language.TODO

Notes for the submission pass

  1. Resolve every TODO by reading the actual licence text and archiving a copy next to this file (LICENSES/<component>.txt).
  2. espeak-ng's GPL-3.0 is the one that can sink the package. Keep Piper strictly an external, user-installed process; do not ship, link, or auto-download it as part of the package payload.
  3. Piper voice models are licensed per dataset (they are not uniformly permissive). List each voice separately, with its own licence, or ship none and have the user download them.
  4. VOICEVOX character terms bind the end user, not the plugin. Documentation must point at them rather than restate them.
← PREVIOUSLicensingNEXT →Voice licences