Grai · Roadmap
Three training runs, in the order in which each one prepares the next. Plus what is already done and what is missing.
Streaming listening, turn-taking decision from prosody, barge-in with correct memory, streaming speech synthesis. Romanian and Russian.
doneForking, a bridge phrase, the result announced separately if it is late.
doneThey survive a reload, a disconnect and a server restart.
doneAgents as records, isolated clients, voice and language per agent.
doneOne line on the client's page, no dependencies.
27.6 KBFLEURS Romanian and Russian, voice-to-voice latency, turn detection — the methodology and the reproduction commands are public.
code and sample: publication decision pendingNot in order of theoretical impact, but in the order in which each step makes the next one possible.
The only thing that closes the 5–9× gap with the competition. The measured fact: on the same 200 FLEURS utterances, our streaming model gets 26.15%, while an open Romanian model, SpeD 110M (UPB, Apache), run offline, locally, on the CPU, gets 6.70%. The offline ceiling exists and is local; we already use it as a second opinion.
The step: distillation — SpeD becomes the teacher of a small student that runs in streaming. It is the next experiment; we give no figure until it has run.
Why first: the turn detector depends on it. Checking that a cut falls on a word boundary is done by transcribing the chunk — with weak recognition, we would build the turn dataset with a crooked ruler.
The smallest of the three, the only publishable one, the only one where we would be first. The reference model covers 23 languages and does not include Romanian, and its authors have announced they are not adding more languages. Romanian would be the twenty-fourth.
The starting point is humblingly low: on Romanian, the existing detector (Smart Turn) gets 4 out of 8 — a coin toss. With n=8 the result is statistically inconclusive: it only shows we have no evidence that it works. The minimum target is to beat chance, on a sample large enough to show it.
Today's speech synthesis engine, Supertonic 3, has had its upstream archived since 23 July 2026, and Voice Builder closed on 31 August 2026. It runs locally, with no upstream development; a replacement is under evaluation. Our own voice has therefore moved from “nice to have” to the critical path.
The technical precedent exists: a Romanian model trained in Cluj used 21 hours of data and twelve hours on a single GPU. But its dataset, SWARA, is non-commercial (CC-BY-NC 4.0 plus a research agreement) — not a path for us. We start from Apache-licensed base models, for Romanian and Russian, plus our own data recorded with consent.
We do not publish a cost figure. The old estimates were made for the August plan, which no longer matches the steps above; we will write down the cost of each step once we have run it. The plan does not depend on startup credit programs — a discount program is a discount, not a precondition.
This list is here because a technical client finds it anyway within the first five minutes. Better to hear it from us.
The real blocker is not compute, it is data. If you have Romanian recordings with clear rights, the first training run happens sooner.