OrOnKol [01]
Speech in, speech out
The real-time layer. Hears the caller over noise, accents and codec loss, and speaks back before the pause becomes awkward.
In production
OrOn Voice Lab
We build the voice, not the wrapper.
Most voice AI is three vendor APIs in a trench coat. We build the layers that decide whether a call feels like a conversation: transport, pronunciation and turn-taking.
Directive
A caller should never have to wonder whether they are talking to a machine, or forgive it for being one.
Tel Aviv · Voice, language and deployment research
01
Models
Nikud [01]
Pronunciation for unwritten vowels
Hebrew and Arabic are written without vowels. Nikud decides how a word is actually said, in context, before a single sound is made.
In productionSicha [01]
Turn-taking and dialogue state
Knows when the caller has finished, when to let itself be interrupted, and how long a silence is allowed to run in that market.
In training02
Research areas
Latency you can feel
We carry audio over UDP, not TCP, so one lost packet never turns into silence.
490ms median replyThe vowels nobody writes
One Hebrew spelling can be three different words. We resolve which one before speaking.
97% word accuracyIt knows who it is talking to
Hebrew and Arabic force a gender choice in every sentence. We never guess.
No wrong addressLevantine is not Gulf
Deployed per dialect and register, so the agent sounds local instead of translated.
32+ localesInterrupt it like a person
Barge-in, endpointing and silence tolerance tuned per market, not set once globally.
Barge-in supportedThe audio never leaves
The whole stack runs on your own hardware when the data cannot cross a border.
100% on-premise03
Findings
Transport
490ms median
Language
32+ locales
Measurement
97% accuracy
Field notes
From deployments
Why we run WebRTC, not WebSockets
One lost packet on TCP becomes a stall. On UDP it becomes nothing.
Nikud, gender and cultural fit
What it takes for an agent to sound native in Hebrew and Arabic, not translated.
The numbers, and how we got them
Measured on production telephony, noisy and accented, against your own baseline.
What broke in production
The post-mortems, not the case studies. Written by the people who shipped it.
04
How we work
↑
Measured, not claimed
Every number on this site is measured on production telephony and re-measured against your own baseline before you sign anything.
→
Native, not translated
A locale is dialect, register, gender and pacing. Translating the words and stopping there is what makes an agent sound foreign.
↓
Yours, not ours
Cloud, your VPC, or fully on-premise. Where the data cannot leave, the whole stack runs inside your perimeter.
←
Explainable, not magic
Every turn is traceable to the stage that produced it. A bad answer is something you can point at and fix.
Next
Bring us the call
Bring us the call
nobody else could handle.
Tell us the market, the languages and the telephony. We come back within one business day with a scoped pilot.