Grai · Voice cloning
The method exists: its author declares six to thirty minutes per voice, on a CUDA GPU. We have not yet run it on a Romanian voice, and we have not measured similarity. It is not the kind of cloning that gets sold, and it comes with obligations most vendors do not put in the contract.
The market sells zero-shot cloning: give five seconds of recording, get the voice. We do not do that, and the reason is that everything doing zero-shot on Romanian is non-commercial — not by accident, but because almost all of those models are trained on the same dataset with a restrictive license.
The method we chose: gradient-based inverse optimization. We have identified it and analyzed it in code; we have not yet run it on a Romanian voice. The speech synthesis model stays frozen, and the voice's style parameters are optimized until they match the target voice. The result is not a new model — it is a file of a few kilobytes that loads into the existing engine.
The method captures speaker identity, not emotional expressiveness. The voice sounds like the person; it does not reproduce the way they get excited or hesitate.
For the real use case — "our receptionist's voice" — that should be enough: the voice does not need to be generated in five seconds, it needs to be generated once, well. How well, we will know after we run it and measure similarity.
We checked the license of every cloning model that supports Romanian, on its own page. The result is a pattern, not a coincidence:
| Model | Romanian | Cloning | License | Commercial |
|---|---|---|---|---|
| OmniVoice | yes | zero-shot | non-commercial | no |
| XTTS-v2 | yes | zero-shot | Coqui Public Model License | no |
| F5-TTS and its Romanian derivatives | yes | zero-shot | inherited from the training set | no |
| Supertonic 3 + style optimizer | yes | through optimization, 6–30 min (declared by the author, CUDA GPU) | OpenRAIL-M on the model, MIT on the optimizer | yes, with obligations |
The cause of the pattern is not the language. It is a single dataset on which almost all zero-shot cloning models were trained. The claim "everything with Romanian and cloning is non-commercial" is almost true — and false in exactly the case we use.
These are not legal footnotes. They are obligations that propagate to every client we install the system for, and we write them here because almost nobody does.
Only Supertonic 3's sample code is MIT. The weights are OpenRAIL-M — commercial use is allowed, but with use restrictions that must be passed on to every client. It is not a blocker. It is a real contract clause.
Written consent from the person, limited purpose, deletion on request. In the Republic of Moldova, Law No. 195/2024 on the protection of personal data has been in force since 23 August 2026 and treats an identifiable voice as a special category of data.
The engine's maker sold a separate voice-building service of its own (Voice Builder), under its own terms; the service closed on 31 August 2026. A style produced locally, through optimization over OpenRAIL-M weights, from a client's recording, is in our reading a derivative work governed by OpenRAIL-M. The reading is reasonable, but it has not been checked by a lawyer. It is on the list for before the first contract that mentions cloning.
Not even at the request of a client who says they have the right. We ask for the speaker's written consent, not a statement from whoever places the order.
Until we run the method on a Romanian voice and settle the contracts, the system ships ten existing voices. The same voice speaks both Romanian and Russian — switching happens mid-conversation, without reconnecting.
The speech synthesis model has 99 million parameters and runs on the CPU, with no accelerator. The model covers 31 languages, of which Romanian and Russian are the ones we use. The Supertonic project was archived upstream on 23 July 2026: it works locally, with no further upstream development; a replacement is under evaluation.
Tell us which voice, what for, and with what consent. We will tell you honestly whether it is possible today, what obligations come with it, and what remains to be settled before signing.