Licensing
What you may ship, and what shipping it obliges. Read this before release, not after.
None of it applies to the plugin itself — Assets/OfflineAINPC/ is proprietary and yours to use under the Asset Store licence. Everything below is about the things the plugin talks to, which you install and may choose to redistribute.
The one that can bite: espeak-ng
Piper's phonemiser links espeak-ng, which is GPL-3.0. Piper's own MIT licence does not change that: a binary distribution of piper.exe carries espeak-ng's terms with it.
The plugin is not affected. It starts Piper as a separate process and talks to it over a pipe; it does not link, load or contain any espeak-ng or Piper code, and nothing under Assets/OfflineAINPC/ is a copy of either. A build check fails if that ever changes.
Your game may be affected, if you ship Piper inside your build. That is allowed and normal. It obliges three things:
- Ship the licence text, naming espeak-ng as the component it covers.
- Offer the corresponding source — a URL to the exact upstream release you shipped, kept available for the period the licence requires. Pin a release tag; "latest master" is not corresponding source.
- Add no further restrictions. Your EULA must not forbid, for that component, what the GPL permits.
Your own game stays proprietary. Launching a process does not make your code a derivative work any more than calling a compiler does.
The escape hatch: do not ship Piper. Point the plugin at VOICEVOX, at a system voice, or at your own engine, and let players who want Piper install it themselves. Nothing in the plugin requires it.
Voice models — provenance, not just the licence line
A voice's dataset licence is not the whole answer. Read its Training line too.
Most Piper voices are fine-tuned from the lessac voice, and the lessac dataset licence (Blizzard 2013, CSTR Edinburgh) excludes commercial speech-synthesis products from its permitted "Research Purposes". A permissive dataset does not clear that: the licence covers the recordings the voice was fine-tuned ON, not the voice it was fine-tuned FROM.
The distinction, in one pair of examples
These two are the same corpus under the same licence, and only one is usable:
| Voice | Dataset | Licence | Training line | Verdict |
|---|---|---|---|---|
en_US-libritts-high | LibriTTS train-clean-360 | CC BY 4.0 | "Trained from scratch on train-clean-360" | usable |
en_US-libritts_r-medium | LibriTTS-R | CC BY 4.0 | "Fine-tuned from English lessac medium" | tainted |
Identical dataset name, identical licence, opposite answers. The Training line decides. Apply that test to any voice you are evaluating.
What this package ships with
| Voice | Licence | Provenance | Where |
|---|---|---|---|
en_US-ljspeech-medium | public domain (LJ Speech) | "Trained from scratch for 1000 epochs" | en standard, en light |
en_US-ljspeech-high | public domain | same, higher quality | en premium |
The default line-up owes no attribution at all, which is deliberate — a requirement nobody notices is how a game ships in breach of a licence it meant to honour.
If you want many distinct voices
en_US-libritts-high carries 904 speakers selected by id, trained from scratch on LibriTTS, CC BY 4.0. It is not in the default presets because CC BY requires attribution and the defaults are deliberately obligation-free. Use it deliberately, and display:
Speech synthesis by Piper (MIT). Multi-speaker English voice "libritts" trained from scratch on LibriTTS train-clean-360 by Heiga Zen et al., licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
Audition speakers with Tools ▸ Offline AI NPC ▸ Speak one line ▸ Audition a LibriTTS speaker....
Every other language
The slots are empty, and the reason is precise: no voice with a traceable commercial provenance was found — the available Piper voices for those languages are fine-tuned from the research-only base. That is a supply problem, not a plugin limitation. Supply your own voice or commission one; see Languages.
Dated copies of every card read, with the full reasoning: LICENSES/VOICE_LICENCES.md.
VOICEVOX character voices carry per-character terms that bind the end user, not you. Surface those terms rather than absorb them — a link in your credits, not a line in your EULA.
Does espeak-ng's GPL reach the audio it produces?
No, and this is worth stating plainly because buyers ask.
The FSF's position is that a program's output is not covered by the copyright on the program's code. Synthesised speech is output; it is not a derivative work of espeak-ng, and you owe nothing on the audio your game plays.
What the GPL does reach is distribution of the binary. That is the boundary this plugin is built around, and running Piper as a separate process — which it already does — is the recognised way to keep copyleft away from your own code. The obligations in the section above apply to shipping espeak-ng.dll, never to the wav files it helped make.
Model weights
The default catalogue is Apache-2.0 (the Qwen3 family), chosen because it is the cleanest licence to redistribute under and needs no visible attribution.
Some other families do require attribution — Llama's "Built with Llama" is the usual one. The manifest has a field for it and the inspector surfaces it as a finding, because a required attribution nobody notices is how a game ships in breach of a licence it fully intended to honour. A line in your documentation does not satisfy it; it has to be visible in the product.
Weights are never bundled with this package. It describes a download; it is not one.
What the plugin itself contains
No DRM, no licence check, no phone-home, no registration, no time limit, no watermark, and no telemetry of any kind — not even anonymous. "No data ever leaves the machine" is the product's core promise, and a usage counter would make it a lie.
One practical consequence, stated plainly: nothing reports back, so nobody can tell how the plugin is being used. That is the intended trade.
Full third-party inventory: LICENSES/THIRD_PARTY_NOTICES.md.