Purposeful script
The sentences mirror real service calls, so the voice learns the exact music it will need — questions that sound like questions, goodbyes that sound warm.
Omniat Ai · Giving the assistant a voice
It starts with a real person.
When you call our assistant, you hear a warm, steady voice in Gulf Arabic, English or Hindi. It isn't an off-the-shelf voice picked from a catalogue — we build each one ourselves, and every one of them begins with a real speaker at a real microphone.
Why go to this trouble
A voice is the first thing a caller notices.It should sound natural, stay the same from the first word of a call to the last, and say local names the way people here say them. That is hard to get from a generic voice.
The pipeline, in five beats
weeks of work per language
01 · RecordA real speakerA few hundred purposeful sentences, in short sessions
02 · CleanEvery clip checkedLevelled, de-noised, and verified by machine listening
03 · TeachThe model learns herFrom how the language sounds to how she sounds
04 · ChoosePeople pick, blindListeners score versions over a simulated phone line
05 · TunePronunciation, by earBrand and place names corrected until a person approves
Five stages, and a person at both ends: a real speaker goes in, and real listeners decide what comes out. The machinery in the middle exists to carry her voice, not to replace it.
01 · The recording
A few hundred sentences, none of them random.
For each language, a speaker sits down with a microphone and reads carefully chosen sentences — the kind of things the assistant actually says on a customer service call: greetings, questions, confirmations, polite goodbyes. The work takes weeks, done in short sessions so the voice stays fresh and consistent from the first clip to the last.
Arabic, fully marked
Every Arabic sentence is written with its full vowel markings, so the speaker reads each word exactly as it should sound — nothing left to guesswork.
Short sessions
Voices tire. Recording in short sessions over weeks keeps the tone steady across the whole set, which is what keeps it steady on every call.
02 · The clean-up
Real recordings are never perfect.
Room noise, the odd stumble, one sentence louder than the next. Every clip is cleaned and brought to the same volume — and then checked by a listener that never gets bored: a speech recogniser plays each clip back and compares what it heard with the sentence that was supposed to be read. Clips where the speaker lost her place, or the audio is unclear, are set aside. Only clips that pass both the ear and the machine make it into training.
03 · The training
From how the language sounds to how she sounds.
We don't start from nothing. Training begins with a voice model that already knows the language in general — its rhythm, its vowels, how sentences rise and fall — and continues on our speaker's recordings until the model takes on her voice, her pace and her tone.
Zero-shot cloning, as a helper
Sometimes the voice should learn from more material than one person can comfortably record. From a short sample of our speaker, a cloning model can read new sentences in her voice without being trained on her first — extra variety for the training set.
Added to, never instead of
Cloned sentences go through the same quality checks as the real recordings, and they join the real recordings rather than replace them. The speaker stays the source of truth.
04 · The choice
people choose, not a numberThe version that scores best on paper isn't always the one that sounds best.
Training produces several versions of the voice along the way — so the final choice belongs to people, not metrics.
Same lines, hidden labels
Listeners hear the same sentences in different versions of the voice, without knowing which is which.
Heard as callers hear it
Every sample is processed to sound like a phone line, because that is how it will actually be heard — a voice that shines in the studio can dull on the wire.
Highest human score ships
Listeners rate each version, and the one people rate highest is the one that goes live.
05 · The tuning
Some words still need help.
Brand names, local place names, words spelled one way and said another. For these we keep a pronunciation guide the assistant checks before it speaks — the same guide the reply polisher applies on every call → Changes to it are made by people who listen to the result: we don't accept a correction until someone has heard it said properly.
What all this buys
Natural
A voice built from a real person keeps a person's warmth — callers relax into it instead of adjusting to it.
Consistent
The same voice, the same steadiness, from the first word of a call to the last — and from the first call to the ten-thousandth.
Local
Names, places and terms pronounced the way people here say them — the detail a generic catalogue voice gets wrong first.
Get in touch
Hear her for yourself.
The voice on the line was built exactly the way this page describes. Call Omniat on a live number and listen for the difference a real speaker makes.