rhasspy

/

piper-voices

like 640

Follow

Rhasspy 315

All three Polish medium voices report a CC0 dataset but a Lessac-derived training base:

The Lessac corpus was distributed under Blizzard Challenge terms, which as I understand them restrict use to non-commercial research. This leaves the status of the resulting Polish voices ambiguous: the speech data is public domain, but the base model is not.

This was raised in general terms in rhasspy/piper#271, but no conclusion was reached there, and it does not cover the Polish voices specifically.

Questions

1. Are model weights fine-tuned from the lessac base considered derivative works of the Lessac corpus for licensing purposes?
2. If they are, does that restriction extend to the generated audio, or only to redistribution of the weights?
3. Is there a recommended Polish voice that does not carry this lineage? I noticed pl_PL-bass was added more recently — is it trained from a different base?
4. Would it be possible to add an explicit Base model licence: field to MODEL_CARD, alongside the existing dataset licence? The dataset licence alone is currently the field most users read, and it is not sufficient to determine whether a voice can be used commercially.

Context

I am building a Polish language-learning application. It is a private project today, so nothing here blocks me — but I may publish it commercially, and I would rather establish the position before building on it. I suspect others hit the same question, since the "CC0 dataset, restricted base" pattern appears across many non-English voices, not just Polish.