OFFLINE AI NPCVoice licences and provenance
Voice models have separate terms from the plugin and speech engine. Check the model publisher's licence, required notices and model card for the version you distribute. Training-data and fine-tune-base records provide additional context; a permissive weight/dataset label does not describe every recording or recorded person's rights.
Qwen3-TTS models and preset voices
The catalog uses Qwen3-TTS CustomVoice GGUF models and a shared codec. The official checkpoints and the Serveurperso GGUF repository advertise Apache-2.0. Nine preset timbres are embedded in the model, not separate voice downloads. The technical report describes training scale, languages and stages, but not a complete recording-permission inventory for these timbres. Keep the required redistribution notices; see Licensing.
Piper: recorded dataset and fine-tune base
Archived cards are in voice-cards-2026-07-30/ and voice-cards-2026-09-23/. The ljspeech-medium, dmitri and sharvard records below also use Piper voice repository revision c10ece1aade47bb51c153c893d14e5bf8e5b7117.
| Voice | Recorded dataset terms / training | Redistribution notes |
|---|---|---|
| en_US-ljspeech-medium / high | Public-domain LJ Speech; cards record training from scratch. | Engine/runtime notices still apply. |
| en_US-libritts-high | CC BY 4.0 LibriTTS train-clean-360; from scratch; 904 speakers. | Retain applicable attribution and notices. |
| en_US-libritts_r-medium | Related LibriTTS-R corpus, CC BY 4.0; card records a lessac fine-tune base. LibriTTS-R is distinct from the original LibriTTS corpus. | Check the exact voice publisher's terms and recorded fine-tune-base terms. |
| ru_RU-dmitri-medium / es_ES-davefx-medium | Cards label their datasets CC0 and identify a lessac fine-tune base. | Check the exact voice publisher's terms and recorded fine-tune-base terms. |
| es_ES-sharvard-medium | Dataset CC BY 3.0, two speakers; card identifies a lessac fine-tune base. | Check the exact voice publisher's terms and recorded fine-tune-base terms; retain applicable attribution and notices. |
| en_US-amy-medium | Dataset URL leads to a repository-wide CC BY-SA label; the voice has no statement of its own; lessac base recorded. | Check the per-voice publisher's terms and recorded fine-tune-base terms. |
| en_US-hfc_female-medium | NonCommercial dataset terms; lessac base recorded. | Check the stated NonCommercial scope and the exact weight publisher's terms before commercial distribution. |
| zh_CN-chaowen-medium | Card records a Xiao Ya fine-tune base; archived Xiao Ya card names NonCommercial BZNSYP. | Check the stated NonCommercial scope and the exact weight publisher's terms before commercial distribution. |
The lessac dataset licence grants research-only use and excludes commercial use of those materials. A permissive dataset used for fine-tuning does not establish permission to use the recorded base. Check the exact voice publisher's terms and recorded fine-tune-base terms before redistribution; dataset terms do not automatically determine the terms of every derived weight file.
Dataset labels also do not establish every weight publisher's terms. Preserve the exact card, model revision/hash and required credits. LJ Speech's public-domain recording claim is useful evidence; it is not a waiver of Piper/espeak/ONNX notices or a statement that the whole speech stack is obligation-free.
Redistributing a voice
For each voice you distribute, retain its model version/hash, source URL, model card and required notices. Check the engine/build, runtime libraries, preprocessing and dictionaries, and any reference recordings separately. Review the terms again when changing a model or voice.