Where benchmark results live once they exist.
Leaderboards and accuracy comparisons are what technical evaluators look for, and showing real fundamental innovation is the point of having the section.
Architecture, in the meantime
What the engine actually is, end to end. Every stage below streams; nothing waits for the stage before it to finish.
One pipeline serves both, with zero code forks.
Alebex Voice Engine — the stream, in order.
No benchmark results are published here yet.
Nothing is claimed here that is not measured. Until figures exist that Alebex will stand behind in public, this section carries architecture rather than scores.
The engine’s five published response timings.
These are the engine’s published response timings, written the same way on every page that carries them.
| Measurement | Value | Note |
|---|---|---|
| Natural human turn-taking | 200 ms | — |
| Typical single-LLM voice AI | 1,500 ms | Waits for the full generation before speaking |
| Alebex, first word | 500 ms | — |
| Alebex, full answer | 1,500 ms | — |
| Anti-hallucination verdict | under 50 ms | Runs alongside synthesis, not in front of it |
A benchmark travels with its method, its samples and whoever ran it. These carry none of that, because they are not one — they are what the runtime is built to do, and they are described on the pages that describe it.
Each of these is a piece of the engine that had to be built rather than assembled, and each is substantial enough to carry its own write-up.
Four rules, applied before anything is published on this page.
Try the live demo
Drop in your website. Alebex reads it, learns what you do, and calls you in under 30 seconds — fully trained, zero setup.