Licensing basis for lessac-derived voices given the Blizzard 2013 corpus restriction
Hello, and thanks for piper-voices. We use several of these voices on-device in a commercial reading app, and we are doing a licensing review before release. I have a question about ancestry that I could not resolve from the cards alone, and I am asking in good faith rather than to make a claim.
Four of the voices we use state they are fine-tuned from the U.S. English lessac medium voice:
en_US-libritts_r-medium (card licence: mit)
en_GB-vctk-medium (CC BY 4.0)
fr_FR-upmc-medium (CC BY-SA 4.0)
de_DE-thorsten-medium (mit)
The lessac card states License: mit and "Trained from scratch", with Lessac Blizzard 2013 as its dataset. That dataset's licence excludes "using the Materials for any commercial purpose, including the development, marketing, commercialisation, sale or licencing of voice synthesis or speech recognition products or services", and states the recordings are licensed "exclusively for those purposes associated with research and exploration of speech technologies".
https://www.cstr.ed.ac.uk/projects/blizzard/2013/lessac_blizzard2013/license.html
My question is narrow: what is the basis for the permissive licences on these cards? Specifically, either of these would answer it completely:
Separate permission was obtained from Lessac Technologies or CSTR covering downstream commercial use; or
The project's position is that model weights are not a derivative work of the training corpus, so the corpus terms do not carry forward.
Both are reasonable positions. I am not asking anyone for legal advice, only for the project's own understanding, in writing, so we can record it.
If there is an existing statement on this I have missed, a link is more than enough. And if the answer is that the question is genuinely open, that is useful to know too, since it would apply to most commercial Piper deployments rather than just ours.
Thanks for the work either way.
Sure, happy to answer. The position of this project is actually different than the two you listed: the license recorded in the model card is only for the dataset directly used for training. Whether or not the the model weights are considered derivative works is left open to the user's discretion. Piper itself and the included voices are intended for personal and research use, but commercial use is not excluded.
Apologies for the non-committal answer, but I've never heard one way or another from a legal authority. I have plans to retrain many of the existing voices with en_US-libritts_r-medium as a base, and all new voices going forward are using it so we can avoid this question entirely in the future.
Thank you for the clear answer, it helps a lot, and sorry for the late reply - only seeing this now.
Two short follow-ups, and then I will leave you in peace.
Our app downloads these voice files to our users' phones, so our review has to cover redistributing the models, not only using them to generate audio. So we understand your answer correctly: does the project hold, or know of, any permission from Lessac Technologies or CSTR that covers redistributing the lessac-derived voices? Or is it right to read the MIT licence on the collection as covering the project's own work, with no statement either way about rights that come from the Blizzard 2013 recordings?
On retraining: the
en_US-libritts_r-mediumcard says it was fine-tuned from lessac medium as well. Will the new base be retrained from scratch first, so the voices built on it no longer descend from lessac? If there is a timeline, or a first voice to test, we would gladly switch and report back.
Thanks again for Piper and for taking the time.
No problem, happy to help!
- The Piper project itself has no statement about the rights coming from the Lessac recordings; users would need to enter a separate agreement with them about redistribution rights.
- I forgot to update the model card for the new LibriTTS-R voice. It has been retrained from scratch, and will serve as the base voice from now on. If there's a particular voice you're interested in being retrained soon, I can bump it up in the queue 🙂
Thank you 😀 that is exactly what I needed, and the retrained LibriTTS-R voice is great news.
Two things, if you do not mind.
Yes please, I would gladly take you up on the offer! The three we would most like to see retrained are fr_FR, de_DE and en_GB, in that order. Every current voice in those languages is fine-tuned from lessac, so there is no clean option for them at all today, and French is the one our own users ask for most. Any single-speaker or multi-speaker voice in those languages would help.
So we pick up the right file: the retrained en_US-libritts_r-medium is the one currently in the repository, and the older lessac-derived weights are gone from main? We mirror the sherpa-onnx packaged copies, which look like they predate the retraining, so we want to be sure of what we are repackaging.
Happy to test anything early and report back on how it sounds on phones. Thanks again for answering so clearly on the Lessac question!
Can you add es_ES-sharvard to the priority list? Thanks!
Small change of plans: I'm switching to a single speaker from HiFi-TTS for the base model. While LibriTTS-R would be ideal, it would take months to train even with the many resources I have access to. I should be able to finish training HiFi-TTS this week or next.
thanks for retraining voices on a clean base! Could you please add ru_RU-dmitri to the queue? Their datasets are CC0, so a version without lessac ancestry would give Russian its first commercially clean voice.
I’d also +1 es_ES-davefx / sharvard. Happy to test and report back. Thank you!