Greets
“Assalomu alaykum! Eshitaman, nima buyurtma qilasiz?”
Hello! I’m listening, what would you like to order?
0:04
An AI phone operator for restaurants and teahouses in Uzbekistan
The phone rings. The order is ready.
NativeAI answers a company’s own phone number, listens in Uzbek and Russian, asks for every detail the caller left out, reads the order back and sends it to the kitchen, or books a room for the day.
Uzbek · Russian · both in one sentence · How a call goes
The problem
Orders in Uzbekistan still come by phone
Restaurants, delivery kitchens and teahouses take orders and bookings on the phone, and every missed or misheard call is lost money. Callers speak the way people speak in Tashkent, and a phone line is not a studio.
- 01
Callers mix languages
A caller in Tashkent says the dish in Uzbek, the size in Russian, and corrects the count halfway through the sentence.
- 02
Phone audio is narrow
A call arrives at 8 kHz. Most published Uzbek speech results are measured on clean 16 kHz read speech, not on phone calls.
- 03
A guess is a wrong order
“Two lavash, one of them spicy” says nothing about the second. A system that fills the gap delivers the wrong food. NativeAI asks.
Listen
How the operator sounds
Press play. This is NativeAI’s own Uzbek voice, saying sentences the system really says on a call. The English is a translation.
Asks what was not said
“Xo‘p, ikkita lavash: bittasi achchiq. Ikkinchisi-chi, achchiq bo‘lsinmi yoki oddiy?”
Right, two lavash: one spicy. And the second, spicy or plain?
0:05
Reads it back
“Demak, ikkita lavash: bittasi achchiq; bittasi oddiy. Yana bitta yarim litrlik Coca-Cola. Manzilingiz: Chilonzor o‘n to‘rt. Hammasi to‘g‘rimi?”
So: two lavash, one spicy, one plain. And a half-litre Coca-Cola. Your address: Chilonzor 14. Is that all correct?
0:09
Says goodbye
“Rahmat, buyurtmangiz qabul qilindi! Yoqimli ishtaha!”
Thank you, your order is placed! Enjoy your meal!
0:04
A synthetic voice (navoiy-tts) in testing, made on our own GPU, not cloned from anyone. Its quality on real calls has not been measured yet.
Correctness first
The model proposes. The code decides.
A language model understands the caller and proposes a change. Only checked, deterministic code changes the order, its prices or its state. Every proposal takes this path.
- Caller’s words
- Model reads
- Schema check
- Rules
- Menu and price
- Order state
- Audit log
Model’s proposal“Second lavash: plain”
RulesStopped: the caller never said so. Instead of a guess, a question.
No default is inventedA value is filled in only when the business itself set a default for it. Otherwise the caller is asked.
Each company apartThe database separates companies with row-level security: one never sees another’s data.
Everything in the logEvery change goes into an append-only log: who, when and what.
How it works
From the call to the kitchen
- 01
On your own number
No new number. A company with an IP PBX sends calls over SIP. A plain mobile SIM forwards its calls to a city number that reaches us.
- 02
Greets and listens
Takes orders in working hours. Outside them, it notes the caller’s number for a call back.
- 03
Asks what is missing
Size, spicy or plain, a sauce: every detail a dish needs is asked. Anything off the menu is said so.
- 04
Address and payment
The exact street and house. Cash or card, as the restaurant works.
- 05
Reads it back
The whole order is read to the caller. Nothing is placed before the caller says yes.
- 06
Sends it on
To the staff console, to the till (iiko or r_keeper), or to the company’s own CRM. A card payment link goes by SMS.
Technology
Our own speech models, on our own GPUs
The parts that hear and speak run on hardware we control. The language model sits behind an interface we can switch, and checked code owns every decision.
Uzbek speech recognition
GigaAM, run on our GPU. On public read speech through a simulated phone line (FLEURS and USC test sets), it made the fewest word errors of the seven Uzbek models we measured: 10.9% on FLEURS. Real calls are not measured yet.
Our own Uzbek voice
navoiy-tts for Uzbek and CosyVoice for Russian, both served by our own workers on our own GPU.
A language model behind an interface
Claude by Anthropic reads what the caller means and proposes. Recognition, the model and the voice each sit behind a provider interface with failover; when a part fails, the call goes to the staff, never to a guess.
A deterministic order engine
Schema checks, business rules, the menu’s prices and the order’s state are code. The model never changes an order, a price or a payment itself.
The company’s own line
Asterisk with AudioSocket. Calls come over SIP from an IP PBX, or from a mobile SIM forwarded to a city number.
Each company sealed off
PostgreSQL with row-level security per company, an append-only audit log, a staff console and a Telegram bot for staff.
Why GPUs
Every call runs two neural models
Each call needs speech recognition and our voice at the same time, live. Measured on a laptop’s RTX 4060: recognition hears a second of speech in about 0.01 s, in 3.7 GB of GPU memory, while our voice needs 1.2–1.4 s to make a second of speech, still slower than real time, its first audio after 2.4–2.8 s, each speaking worker taking 3.7–5.6 GB. By our estimate, a lunch peak of twenty calls at once needs several such workers: production needs data-centre GPUs.
- Recognition: time to hear 1 s of speech
- 0.01s
- Voice: time to make 1 s of speech
- 1.2–1.4s
- GPU memory per voice worker
- 3.7–5.6GB
Measured by us on an RTX 4060 Laptop GPU, 2026. Recognition: public read speech through a simulated phone line, not real calls.
Contact
Let NativeAI take the next call.
NativeAI is an early-stage team in Tashkent, Uzbekistan, founded by Elshod. Write to us by email or on Telegram; companies that want to connect fill in a short form there, in Uzbek.