A voice built for the way we speak.

No. There is no public demo, 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, 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%. Protocol: 200 utterances per language from FLEURS, seed 20260831, streaming decoding. How good it is: it is not good. Our own conclusion, written in the comparison table, is that on Romanian accuracy we fall behind commercial systems, by a factor of five to nine. Romanian is almost three times harder to recognise than Russian with the same configuration, which is exactly 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 almost as well. Anyone who wants a faster voice conversation needs to attack the language model — by using a smaller model, by streaming the response or by shortening the replies. Here we must also be honest with ourselves: the segment figures published on the site do not add up to the total published on the same screen, and the difference is about half a second. The breakdown is the one we trust; the aggregated total must be recalculated before it is cited.

Because in a cascade you can change each component separately and keep recognition and synthesis local, on the processor, without GPU. The price is chaining latency and the fact that the language model, which is the largest part of the delay, is outside the machine. Native duplex is a direction, not something we have — it is on the list of ‘does not exist yet’.

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

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

Because otherwise there would be no point in publishing them at all. A table in which you always come first is a marketing table. Ours marks third-party rows as declared and only our own as measured, and the conclusion written under it says that we are losing. It must still be made clear what no outsider 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.