Forty years of better screens.Your staff have their hands full.
Every improvement to restaurant software has been a better screen — bigger, faster, nicer to tap. Meanwhile the captain is carrying four plates and the chef has both hands in a tandoor. Chotu is the part that talks. You say “Suno Chotu” and he takes the order. Then he does the thing nothing else does: he speaks first.
Why “Suno Chotu”, and not just “Chotu”
In half the kitchens in India, chotu is somebody’s actual name — the young one who runs errands, the person you shout for across a loud room. That is exactly why the name was chosen, and exactly why it could not be the wake word on its own: a bare “Chotu” collided with a real person being called, all shift.
So it is two words. Suno Chotu — listen, Chotu. An imperative and a name, which is how you summon anyone across a kitchen anyway. It is also acoustically far easier to pick out of clatter than one short word, and it needs far fewer negatives to train.
He hears it on the device. A 1.1 MB model, running on the phone in your apron, against a rolling one-second window. Nothing leaves the handset until it fires.
One ticket
From the floor to the tandoor,
Dinner service, table 4. Two butter naan and a paneer tikka. Watch who gets spoken to, and when — that is the part no other voice feature does. Some of the lines you can hear in his real voice.
The captain
On the floor
Ramesh
At the tandoor
The head chef
On the pass
- 20:41
The captainChotu
Suno Chotu — table 4. Do butter naan, ek paneer tikka.
Hands full, order pad nowhere. The wake phrase is two words on purpose: “Chotu” alone is a real name in half the kitchens in India, and a bare wake word collided with it constantly.
Working today - 20:41
ChotuThe captain
Theek hai.
Punched, KOT fired, exactly as if somebody had tapped it. Voice returns an action descriptor and the same handler a button uses applies it — the bill never depends on speech recognition being right.
Working today - 20:41
ChotuRamesh
Butter naan banao, table 4.
Here is the turn nothing else does. Chotu speaks first, into one person’s ear, chosen because Ramesh is on the tandoor station and his session says free.
Built, not switched on - 20:42
RameshChotu
Shuru kar diya.
No wake word. Chotu asked, so the next thing Ramesh says is the answer — a 90-second reply window in the kitchen, because a chef with his hands in a tandoor does not answer in eight.
Built, not switched on - 20:49
ChotuRamesh
Ready hai kya?
Asked when the dish could plausibly be done, not when it was punched. Asking at punch time would only teach the kitchen to ignore him.
Built, not switched on - 20:50
RameshChotu
Ek jal gayi.
One burnt. Because the naan carries a recipe, this is meant to deduct the flour, the butter and the ghee — not write a note nobody reads. Nobody logs one burnt naan today, because logging it means walking to a screen.
Designed, not built - 20:51
ChotuRamesh
— silence —
The second naan goes unanswered. He re-asks on doubling gaps — 60 seconds, then 120, then 240 — so three reminders spread across the rest of the cook time instead of landing inside one minute.
Built, not switched on - 20:56
ChotuThe head chef
Jawab nahi mila.
Four asks, no answer, so it goes to a human. The status stays explicitly unknown — he will not guess “ready”. The comment in the code reads: going quiet is the one outcome never allowed.
Built, not switched on - 20:57
The captainChotu
Chotu, table 4 kitna time?
The walk to the pass to ask this is one of the most repeated wasted trips in a service. Everything needed to answer it is already in the system.
Built, not switched on - 20:57
ChotuThe captain
Butter naan 4 minute mein.
And it rounds up. Promising early is worse than promising late, because the guest starts counting from whatever number the captain repeats to them.
Built, not switched on - 21:00
ChotuThe captain
Table 4 ka butter naan ready hai.
Back to the floor, in the ear, without the captain looking at anything. Routine lines wait while your own mic hears you talking — but this one interrupts, because a swallowed “table 4 ready” is worse than a late one.
Built, not switched on
The voice you hear is the one the product ships — the same rendered lines that sit in the Android app, not a recording made for this page. Lines that name a dish, a table or a number of minutes are spoken live by the product and have no file to play here.
What he is not allowed to do
A voice that can write to a bill is a voice that can ruin one. Every rule below is in the code because something went wrong in a real service, or because somebody worked out that it would.
One wake, one command
There used to be a 22-second listening window. A pilot at a real store caught a guest’s own speech being executed as orders — and a voiceprint cannot fix that, because a waiter talking to a guest is still the waiter’s voice. Every command is addressed individually now.
He can only change what he asked about
A chef saying “ho gaya” into the air does nothing. The reply has to match an open question, and it has to come from the person that question was put to.
“Ready nahi” can never mean ready
The grammar carries an explicit negation guard — nahi, nahin, nai, mat, not — so a denial can never be matched as a confirmation. “Punch” is only ever read as a whole utterance, because “paanch” is the number five.
Money is confirmed out loud
Settling asks twice. “Total ₹840. Cash, UPI ya card?” then “₹840 cash — settle kar doon? Haan ya na.” Negatives always beat affirmatives, so a muttered “nahi” can never be read as a yes.
It knows it is you, on your phone
Your voiceprint is a non-reversible embedding computed on the handset. Raw audio never leaves it. Writes need the wake word, your voice and a matching command; listening stays liberal, because the worst case there is an unneeded whisper.
Nothing bypasses the visible record
Every voice write lands on a screen where a human can see it and undo it. If voice is slow or down the buttons are still there — the worst case is that voice did not help, never that the POS is stuck.
Hinglish is the engineering, not the adjective
The command grammar is around forty intents with hundreds of hand-written phrase variants, matched whole-word, then folded phonetically so accent and transcription drift resolve without listing every spelling.
It reads the number words — ek, do, teen, chaar, paanch — the portion words, aadha and poora, the fillers people actually open with, haan, bhai, boss, and the imperatives they close with, kar do, de do, laga do.
And it will never strip a bare “do” as a verb, because “do” is also the number two. That one line is the difference between two naan and no naan.
Where this actually is
A page like this normally blurs the line between what runs and what is planned. Here is the line, drawn. Ask us on a call and we will tell you what moved since.
Working today, and demo-able
- The “Suno Chotu” wake word, running on-device — a 1.1 MB model, no cloud round trip to hear you
- Taking an order by voice at the counter: add, modify, cancel, apply an offer
- Settling and printing by voice, with the two-step spoken confirmation
- Hinglish speech recognition tuned for Indian kitchens rather than adapted from English
Built and tested, not yet switched on
- The dialog spine — open questions, widening re-asks, escalation to a human
- Directed whisper: one person, one earpiece, priority and deferral
- The kitchen consumer that picks which chef to ask
- The Android app that carries it in an apron pocket, on eight of nine milestones and not yet run on real hardware
Designed, not built
- Waste by voice — “ek jal gayi” deducting the burnt naan’s ingredients through its recipe
- Guest-assist: menu answers whispered to a captain standing at the table
- Station rostering and load-aware routing across a full brigade
- Inventory by voice — raising a purchase order and receiving goods
Questions people actually ask
Does everyone need a headset?
One single-ear boom-arm headset per person, on their own phone. Single ear is not a cost decision — the captain still has to hear the guest and the chef still has to hear the kitchen. Always-on listening keeps the Bluetooth link open, which is what these headsets rate as talk time, so a long shift usually means a spare on charge. The headset button is wired to the same code path as the wake word, so a store that wants push-to-talk instead has it without a new build.
Does it work in a loud kitchen?
That is the environment it is designed for, and it is also the thing we have not yet measured. The wake model is trained with narrowband augmentation because Bluetooth headsets often negotiate an 8 kHz link, and the detector runs conservatively so it would rather miss a wake than fire on a clatter. We would rather tell you that plainly than quote you an accuracy number from a lab.
What happens when the wifi drops?
Hearing you is on-device, so the wake word does not need the network. Open questions live on the server, so a dropped phone loses no state — when it reconnects, Chotu re-states the one question still pending rather than replaying everything you missed. A dead phone’s session goes away on its own and the question escalates rather than being silently swallowed.
Is it recording my restaurant all day?
No. The wake word runs on the device against a rolling one-second window and nothing leaves the phone until it fires. Guest audio is transcribed and discarded, never stored. Voiceprints are non-reversible embeddings computed on the handset — the audio they are made from never leaves it, staff enrol themselves, and they can delete their own.
What if a member of staff has not enrolled?
They still work — being logged in on their own phone is enough to be addressed. Enrolling adds the check that stops someone standing near the counter from issuing a command in your name, which matters more at a roadside counter than in a kitchen: a kitchen is loud, but every voice in it is on your payroll.
When can we actually use it?
The counter half demos today and is running at a pilot store. The kitchen half is written and tested and needs to be switched on against live orders, which is a rollout decision rather than a build one. Ask us on a call where it is this month — we will tell you what is running rather than what is planned.
Take an order without touching anything.
Fifteen minutes. We put your menu in, you say it out loud, and you watch the ticket reach the kitchen.