Voice licences — read, not summarised

Every card below was opened and read on 2026-07-30, and a dated copy saved in voice-cards-2026-07-30/. Cards change; this folder is the record of what was relied on.

Read the finding at the bottom before using any of these.

What each card says

Required attribution strings: none stated on any card. CC-BY-SA and CC-BY-NC-SA both require attribution by their terms, but neither card supplies the string to use.

Two findings, in order of severity

1. hfc_female is NonCommercial and must not be recommended

en_US-hfc_female-medium is CC BY-NC-SA 4.0. A buyer shipping a commercial game with it is in breach. It was listed as "ready out of the box" in this package's own documentation until this was read — that listing was wrong and has been corrected.

It is also the voice this demo project uses for one character. That is a problem for the game, separately from the plugin.

2. Every one of these voices is finetuned from lessac, whose dataset forbids commercial use

The cards for amy, hfc_female, davefx and dmitri all say, verbatim:

Finetuned from U.S. English lessac voice (medium quality).

And the lessac dataset licence (Blizzard Challenge 2013, University of Edinburgh CSTR) says "Research Purposes" excludes

…developing, adapting, amending or otherwise using the Materials for any commercial purpose, including the development, marketing, commercialisation, sale or licencing of voice synthesis or speech recognition products or services…

This is not a conclusion, it is a flag. Whether a finetuned model's weights are a derivative work of the training corpus is a legal question with no settled answer, and it is not one a developer should answer for a shipping product. What is established:

• the dataset those weights were finetuned from prohibits commercial voice-synthesis products;
• a CC0 dataset (davefx, dmitri) does not clear the lessac provenance, because the CC0 applies to the recordings that were finetuned ON, not to the voice that was finetuned FROM.

Consequence for stage 5E's conclusion: it recorded ru_RU-dmitri as "a voice with a licence fit for commercial redistribution". That was based on the dataset line alone and did not account for the provenance line on the same card. The dataset is still CC0; the claim that the voice is commercially clear is no longer supportable without legal advice.

What this package does about it

The shipped preset catalog already leaves every Piper voice slot empty, with unavailableReason explaining that the per-voice licence must be established before shipping. That was the right call and it stays: nothing here promotes any voice into a preset.

The documentation has been corrected — see Documentation/Licensing.md and Documentation/Languages.md. Neither now presents a Piper voice as cleared.

For the buyer, the honest position is: Piper is a fine engine, the voices are a supply problem, and clearing one for a commercial product is work only they can do for their own jurisdiction and risk appetite. The plugin points at the cards; it does not vouch for them.

Resolved 2026-07-30 (stage 5L): two voices with a clean provenance

The problem above has an answer. Two Piper voices are trained from scratch, and their cards were read before anything was built on them:

And the near-miss, which is the more useful record. en_US-libritts_r-medium is the same corpus under the same CC BY 4.0 licence, and its card says:

Fine-tuned from English lessac medium on train-clean-360.

Same dataset name, opposite answer. The Training line decides, not the dataset name. Its card is saved here too, deliberately, as the counter-example.

What shipped

The default presets use ljspeech throughout, because it is public domain and therefore owes no attribution at all — the defaults are deliberately obligation-free. libritts is documented as an opt-in for projects that want 904 distinct voices and will display the CC BY attribution.

Every other language keeps an empty slot with the precise reason: no voice with a traceable commercial provenance was found for it.

Measured cost of the swap

Synthesis is slower than the tainted voices were. Same phrase, same machine:

The new default is roughly twice the synthesis cost of the old one and still comfortably faster than real time. The high voices are at or past real time, which matters for the latency budget: they are appropriate for the premium tier and for opt-in use, not for a machine that is already struggling.