One voice at a time
A floor controller decides who has the line, so nobody hears two voices at once. Speak up to cut in.
ಕನ್ನಡ · हिन्दी · मराठी · ગુજરાતી · தமிழ் · English
Speak Kannada.
Be heard in Hindi.
Convox interprets phone calls live, inside the ride, delivery and service apps that already connect two strangers. Each person speaks their own language and hears the other in theirs. Nobody has to switch.
· 00:00
Convox interpreting
Connecting
· 00:00
Convox interpreting
Connecting
Speakers are synthetic voices; the translation and the interpreter's voice are Convox's real output, timed when generated.
Built on
- Sarvam AI
- Exotel
- Bolna
Saaras recognition · Mayura translation · Bulbul voices
The problem
A Bengaluru auto driver calls his passenger to find the pickup. He speaks Kannada. She speaks Hindi. The ride gets cancelled, or the argument becomes the next viral video. Neither of them should have to give up their language to get a ride.
It keeps happening: The Week, April 2025 · Free Press Journal · Deccan Herald
And it isn't only autos. The same call fails all day:
- The parcelThe delivery agent can't find the address or get the OTP. The delivery fails, and a parcel sent back costs the courier ₹150–250.1
- The 10-minute orderThe rider is at the wrong gate with the clock running. The order arrives late or gets cancelled.
- The repair visitThe technician never understands the actual problem. The visit is wasted and rescheduled.
- The clinic callThe patient explains in Gujarati, the doctor answers in English. The dose is half understood.
- 45 cr+
- Indians live away from where they were born, Census 2011
- 22
- scheduled languages, and a call carries one at a time
- 0
- apps to install. It's the call people already make
How it works
An interpreter that joins the call and stays out of the way.
- 01
Listens
Knows when you've finished a thought. A "haan" or "hmm" doesn't interrupt.
- 02
Understands
Sarvam speech recognition in six Indian languages, detected from your first two seconds.
- 03
Translates
Sarvam Mayura translation. Every number is checked, so 47296 stays 47296.
- 04
Speaks
A natural Bulbul voice in the listener's language, after a soft chime.
Latency, honestly
A beat of quiet after each line.
Convox interprets the way a person does: it waits until you finish a thought, then the other person hears it in their language. That wait is the price of a whole sentence instead of broken fragments. Here is what it costs on real calls.
Speaker
talks · 4.8 s
Convox
1.3 s
Listener
hears the translation
- 0.30 s knowing you've stopped
- 0.66 s translating
- 0.19 s the voice starting
- medians of each step; they don't add up exactly
Typical quiet before the other person hears you
1.3s
Median of 65 lines on 9 live calls. The slowest one line in ten waited 3.9 s or more.
Long sentences start before you finish
7.7s → 4.3s
Whole sentence
Clause by clause
Speech start to first translated word, live Marathi → English. Never splits a number.
The live demo
Watch a call get through.
Two phones, two people, no shared language. Each column is one person's screen, entirely in their language; between them, the quiet before each line is heard. When they hang up, Convox writes down what they agreed, in both languages. This is the real demo screen playing a sample call.
For engineers, every call also has an under-the-hood view: the pipeline split of each line, who has the floor, and every number check.
Why it feels like a real call
Built for how India talks on the phone.
Numbers come out right
OTPs, amounts, flat and phone numbers are checked digit by digit between what was said and what is heard.
Knows your language
Detected automatically. If it can't tell, one keypad press picks it.
No app, any phone
Works on a normal phone call through Exotel. For demos, two phones scan a QR code.
Keypad control
1 replays the last line. 0 turns interpreting off.
Private by default
No recording in production. Transcripts only with consent, in line with the DPDP Act.
Who it's for
Apps whose users already call each other through them.
Ride, delivery and service apps already route every driver and customer call through a masked number. Convox joins that call, so the platform turns it on once and every driver, rider and customer gets it without installing anything.
Ride hailingStart here
Autos, cabs and bike taxis
Driver and passenger on "where are you?". A missed pickup is a cancelled ride.
ಕನ್ನಡ ⇄ हिन्दीParcel delivery
Couriers and e-commerce fleets
Address, landmark and OTP. Every failed call is a failed delivery the courier pays for twice: the retry and the return.
தமிழ் ⇄ हिन्दीQuick commerce and food
Riders at the gate
Ten minutes on the clock and the customer gives directions in a language the rider doesn't speak.
मराठी ⇄ हिन्दीHome services
Technicians and customers
The job is understood before the visit, not after a wasted trip.
ಕನ್ನಡ ⇄ हिन्दीClinics and telehealth
Patients and doctors
Symptoms in the patient's words, and a dose that arrives intact.
ગુજરાતી ⇄ EnglishSupport desks
One desk for every state
No routing calls by language, no transfers, no "please hold for a Tamil speaker".
தமிழ் ⇄ EnglishHow platforms use it
Inside the platform's calls, not another app.
Why not an app
iPhone apps can't hear phone calls at all, and Android locks call audio down further every release. Interpreting also needs both voices, and an app on one phone only has one.
Where it runs
Masked calls already pass through the platform's cloud telephony, such as Exotel, where both voices are. Convox joins the call there. Nothing changes on either phone.
When it switches on
Only when the two people's languages differ (the app already knows from their profiles), when a driver taps a language button, or after a failed call. Every other call is untouched.
The pilot
One city, one hub, four weeks. Compare missed pickups or failed deliveries on mismatched-language calls, with Convox on and off.
Try it with two phones.
Open the demo on a laptop, pick two languages, and scan one QR code with each phone. Then just talk.