A voice built for the way we actually speak.

No. There is no public demonstration, no public API, no telephony integration, no native duplex, and voice cloning is not in the product. All five are listed as missing in our own roadmap. If you need a voice agent that works now, the answer is Kallina, not Grai; if you need an audio file, it is Voice Studio.

On Romanian recognition, a 26.31% word error rate on the default 1120 ms window, with a 95% confidence interval between 24.07 and 28.60; in Russian, 9.30%. The protocol: 200 utterances per language from FLEURS, seed 20260831, streaming decoding. How good is it: it is not good. Our own conclusion, written under the comparison table, is that on accuracy in Romanian we lose to the commercial systems, by a factor of five to nine. Romanian is almost three times harder to recognise than Russian on the same configuration, and that is precisely the problem we are studying.

On the measured breakdown, the language model takes 2180 ms of the path, recognition 92 ms and synthesis 305 ms. Recognition is practically free; synthesis nearly so. Anyone who wants a faster voice conversation has to attack the language model — a smaller model, streaming the response, or shorter responses. Here we have to be honest about ourselves as well: the per-segment figures published on the site do not add up to the total published on the same screen, and the difference is roughly half a second. The breakdown is the part we trust; the aggregate total has to be recomputed before it is quoted.

Because in a cascade you can swap each component separately, and you can keep recognition and synthesis local, on the CPU, without a GPU. The price is the chaining latency, and the fact that the language model — the largest part of the delay — sits off the machine. Native duplex is a direction, not something we have: it is on the "does not exist yet" list.

The roadmap proposes three training runs under 1,200 dollars in total: recognition for Romanian on VoxPopuli CC0, 89 hours, roughly 50 hours of H100, around 500 dollars; an end-of-turn detector for Romanian, roughly 2,000 samples of our own, under 100 dollars; a voice of our own, 500-800 dollars. The third is not blocked by money but by data — and by the fact that the synthesis model we were relying on has been archived by its authors. The target figure for the first run, around 21%, is an extrapolation from the improvements obtained on Greek and Bulgarian, not a measurement.

The method being studied is inverse optimisation over frozen weights, not cloning from a few seconds: 6-30 minutes of material per voice, roughly 3,000 optimisation steps, 2.6 GB of peak memory, with the delivered artefact on the order of kilobytes. But it is not in the product, and it matters to say why that is not merely a technical question: in the Republic of Moldova the voice has been biometric data since 23 August 2026. Without explicit consent and an assignment of rights, the question is not whether it can be done, but that it is not done.

Because otherwise there would be no point publishing them at all. A table in which you always come first is a marketing table. Ours marks the third-party rows as declared and only our own as measured, and the conclusion written beneath it says that we lose. It also has to be said plainly what nobody on the outside can do: the code that produced these measurements is not public and does not accompany the site, so the figures cannot be reproduced by a third party today. Treat them as internal measurements, not as independently verified results.