Call once,
it's booked.
Vakya answers the phone for a clinic, a hotel or any front desk in Hindi, Marathi and English, checks the schedule, reads every detail back, and books only after a spoken yes.
- Hindi
- Marathi
- English
- Code-mixed
Read the whole call.
The real call from the demo, word for word as Vakya heard it: one booking, a second one, and a cancellation, with WhatsApp and Google Calendar updated at each step.
Vakya
नमस्ते Ayush जी, फिर से आपका स्वागत है। आज कैसे मदद करूँ?
Hello Ayush, welcome back. How can I help today?
Caller
मुझे बुधवार शाम छह बजे बुकिंग करनी थी।
I wanted a booking on Wednesday at 6 in the evening.
- Checks Wednesday 6 PM: taken. Finds the nearest free times.
Vakya
बुधवार 11 नवंबर शाम 6 बजे का समय भरा हुआ है। क्या आप शाम 5 बजे या शाम 7 बजे चुनना चाहेंगे?
Wednesday 11 November at 6 PM is taken. Would you like 5 PM or 7 PM?
Caller
हाँ, चलेगा। मुझे सात बजे कर दीजिए booking।
Okay. Book it for 7, please.
Vakya
बुधवार 11 नवंबर शाम 7 बजे, Ayush के नाम से, फ़ोन नंबर के आख़िरी अंक 6 2 9 5. क्या मैं बुक कर दूँ?
Wednesday 11 November, 7 PM, for Ayush, number ending 6 2 9 5. Shall I book it?
Caller
हाँ, कर दीजिए।
Yes, go ahead.
- Booked A107, Wednesday 7 PM. Added to Google Calendar. WhatsApp sent
Vakya
बुकिंग हो गई है। बुकिंग आईडी A107 है, बुधवार 11 नवंबर, शाम 7 बजे। आपका टोकन नंबर 11 है। पुष्टि का संदेश WhatsApp पर भेजा जाएगा।
Booked. Your booking ID is A107, Wednesday 11 November, 7 PM. Your token number is 11. The confirmation is on WhatsApp.
Caller
Alright, you know what… गुरुवार को ग्यारह बजे करना है।
Alright, you know what… I want Thursday at 11.
Vakya
गुरुवार 12 नवंबर सुबह 11 बजे, Ayush के नाम से, फ़ोन नंबर के आख़िरी अंक 6 2 9 5. क्या मैं बुक कर दूँ?
Thursday 12 November, 11 AM, for Ayush, number ending 6 2 9 5. Shall I book it?
Caller
हाँ, चलेगा।
Yes, fine.
- Booked A108, Thursday 11 AM. Added to Google Calendar. WhatsApp sent
Vakya
बुकिंग हो गई है। बुकिंग आईडी A108 है, गुरुवार 12 नवंबर, सुबह 11 बजे। आपका टोकन नंबर 1 है।
Booked. Booking ID A108, Thursday 12 November, 11 AM. Your token number is 1.
Caller
You know what, मुझे वो वाली booking cancel करनी है।
You know what, I want to cancel that booking.
Vakya
कौन सी बुकिंग रद्द करनी है? बुधवार 11 नवंबर शाम 7 बजे (A107) या गुरुवार 12 नवंबर सुबह 11 बजे (A108)?
Which booking should I cancel? Wednesday 7 PM (A107) or Thursday 11 AM (A108)?
Caller
A one zero seven booking cancel करना है मुझे।
I want to cancel A107.
Vakya
आपकी बुकिंग A107 है, बुधवार 11 नवंबर शाम 7 बजे। क्या मैं A107 रद्द कर दूँ?
Your booking A107 is Wednesday 11 November at 7 PM. Shall I cancel A107?
Caller
हाँ, चलेगा, कर दीजिए।
Yes, do it.
- Cancelled A107. Removed from Google Calendar. WhatsApp sent
Vakya
बुकिंग A107, बुधवार 11 नवंबर शाम 7 बजे, रद्द कर दी गई है। पुष्टि का संदेश WhatsApp पर भेजा जाएगा।
Booking A107, Wednesday 11 November at 7 PM, has been cancelled. The confirmation is on WhatsApp.
Caller
Alright, thank you. धन्यवाद। Bye.
Alright, thank you. Bye.
Recorded on a real phone line, 9 November 2026. The clinic and its other patients are sample data; the caller is a member of our team. WhatsApp messages were sent from a test number.
One call, four moves.
The read-back is the product. Phone lines are noisy and digits get misheard, so Vakya says what it understood and waits for a yes.
- 1
It answers first
A caller rings the business's number, a clinic's, a hotel's or a salon's, or gives a missed call and Vakya rings back. The agent speaks first and says it is an AI.
- 2
It listens and answers
Hindi, Marathi, English and the mix of all three. Answers come from the business's own documents and its schedule, not from the model's memory.
- 3
It reads everything back
Day, date, time and phone number are read aloud. The booking is made only after the caller says yes, and a wrong digit is fixed on the spot.
- 4
It leaves a signed receipt
Every clip it speaks gets a signed receipt anyone can check at /verify. A person can watch the call live and take over.
Sunrise Clinic (invented)
Agent
नमस्कार, मी सनराइज क्लिनिकची AI असिस्टंट आहे. सांगा, काय मदत करू?
Hello, I am Sunrise Clinic's AI assistant. How can I help?
Caller
गुरुवारी संध्याकाळी डॉक्टरांची appointment हवी आहे.
I need a doctor's appointment on Thursday evening.
Agent
गुरुवार, ५ नोव्हेंबरला संध्याकाळी ५:३० वाजता वेळ मोकळी आहे. तुमचा फोन नंबर सांगा.
Thursday 5 November at 5:30 pm is free. Please tell me your phone number.
Caller
नऊ आठ सात सहा पाच, चार तीन दोन एक शून्य.
Nine eight seven six five, four three two one zero.
AgentRead-back
मी परत सांगते: गुरुवार, ५ नोव्हेंबर, संध्याकाळी ५:३०, फोन नंबर ९८७६५ ४३२१९. बरोबर आहे का?
Let me read it back: Thursday 5 November, 5:30 pm, phone 98765 43219. Is that right?
CallerCaller corrects one digit
नाही, शेवटचा अंक नऊ नाही, शून्य आहे.
No, the last digit is not nine, it is zero.
Agent
माफ करा. फोन नंबर ९८७६५ ४३२१०. आता बरोबर आहे का?
Sorry. Phone 98765 43210. Is that right now?
CallerSpoken yes
हो, बरोबर.
Yes, correct.
AgentCommitted
बुक केलं आहे: गुरुवार, ५ नोव्हेंबर, संध्याकाळी ५:३०. तुमची सही केलेली पावती तयार आहे.
Booked: Thursday 5 November, 5:30 pm. Your signed receipt is ready.
Written by us to show the read-back. The clinic, the caller and the number are made up; it is not a recording.
What Vakya does.
One front desk that answers the phone, checks the schedule and confirms what it booked.
- vobizgnani.ai
A real phone call, not a browser demo
Callers dial a real number over Vobiz telephony (8 kHz phone audio). Gnani Prisma hears them live, Gnani Timbre answers in Hindi, Marathi or English.
Thinks before it books, and remembers you
It checks real availability, offers the nearest free slots and reads every detail back. It only books on a clear yes. Reschedule or cancel in the same call, and returning callers are greeted by name: “Welcome back, Ayush.”
WhatsApp, the moment it acts
The confirmation lands the moment it books, moves or cancels, while the caller is still on the line.
This demo sends from a WhatsApp test number.
Straight into your Google Calendar
Every booking, move and cancellation lands in the owner's own Google Calendar, so they know the moment someone books.
Watch live, take over
The owner can watch the live transcript of any call and take it over at any moment.
A receipt for every booking
Every booking gets a verifiable receipt: who called, what was said and how it was booked.
Hindi, Marathi, English
Understands Hindi, Marathi, English and code-mixed speech.
Answers from your documents
Upload your FAQ, price list or policy (PDF, TXT, Markdown). Search understands Hindi and Hinglish spellings, and questions it could not answer are listed for you to fill in.
Visual schedules
Set opening hours and slots by sight, not forms.
Callback and audit
Missed-call callback, call replay and a booking audit trail.
1. From your FAQ
Q: Sunday ko clinic khula hai?
A: Sunday ko clinic band rehta hai. Monday se Saturday, subah 9 se shaam 5 baje tak.
2. Caller asks, in Hinglish
“Sunday ko doctor milenge kya?”
3. Spoken answer
“Sunday ko clinic band rehta hai. Aap Monday se Saturday, subah 9 se shaam 5 baje ke beech aa sakte hain.”
Visual schedules
Set your hours by sight.
Click a slot to open or close it. That is the whole setup: no forms. Watch the grid for a few seconds and a call books one.
An illustration with sample data, not a live schedule.
- Open
- Closed
- Booked
- Booked by phone
Waiting for a call…
For every front desk.
Clinics, hotels, salons, restaurants, offices, institutes, service centres and gyms. Pick yours, name it, and it answers.
Clinic Hotel Salon and spa Restaurant Professional office Coaching and admissions Service centre Gym and studio
Built on Gnani
The speech is Gnani's.
Vakya calls Gnani's API directly. Nothing sits in between except the turn manager that decides when to speak and when to confirm.
Live on Gnani since 5 Nov.
Prisma and Timbre have run against the live API since 5 Nov 2026, and the noise report card is measured on Prisma v3. Voice cloning, the realtime stream and a real phone call have not run live yet.
Prisma v3
POST /stt/v3wss://api.vachana.ai/stt/v3/streamHears the caller. REST per utterance today; the realtime WebSocket at 8 kHz is being wired onto the phone path. format=codemix keeps English words in Latin and numbers spelled in the caller's language. keyterms carry the business's staff and service names from its own documents; previous_text carries the caller's last sentence.
Timbre
POST /api/v1/tts/inferenceSpeaks for the front desk. Raw 8 kHz audio goes straight to the phone line, so there is no extra resampling. Reads dates, amounts and phone numbers the way an Indian caller expects.
Voice cloning
POST /api/v1/tts/voice-clone/embeddingsmodel: vachana-vc-v1An owner can give the agent a consented voice. The recording becomes a speaker embedding, and Timbre speaks with it. The consent tick, the label and the receipt stay in front of it.
Reasoning
A hosted language model writes the replies and chooses the tools. Hosted embeddings search the business's documents, with keyword search as the backup. Gnani does the speech; the language model does the thinking. There is no second speech model. When Gnani is busy or down, the caller hears a short Timbre line rendered ahead of time and is offered a callback.
Report card
Measured on Gnani PrismaWhere it breaks, measured
The same 24 invented booking calls are played clean, over an 8 kHz phone line, through street noise and a phone codec, with and without the business's own names as keyterms, and scored on whether the name, day and time all came through.
Measured on Gnani Prisma v3.0. The audio is synthetic speech and the noise is added by software: it shows the shape, not what a real caller will score.
| Line condition | No keyterms | With keyterms |
|---|---|---|
| Clean | 100% | 100% |
| Phone line, 8 kHz | 100% | 100% |
| Phone codec, G.711 | 100% | 100% |
| Street noise, 5 dB | 100% | 100% |
| Street noise, 0 dB | 96% | 100% |
| Phone line and noise | 95% | 100% |
The audio is synthetic and the noise is added by software: repeat on real recordings before relying on a number. Caller Theatre: 9 of 9 scenarios pass, on a scripted model.
Built to be checked.
A voice that sounds human on a phone line needs a paper trail. These are the parts that exist today.
It says it is an AI
Every call opens with a spoken disclosure, before anything else the caller hears.
Cloned voices need consent
The owner ticks a consent statement before a voice is cloned. The record is stored with the voice, and withdrawing it deletes both.
A label and a content manifest
Every clip has an AI-generated label and a C2PA content manifest in the file, so a copy can be traced back to its receipt. It is evidence, not proof: a re-encoded or edited clip no longer matches.
A signed receipt
Each clip gets an Ed25519 receipt. Anyone can drop a clip on /verify, no login, and see whether Vakya made it.
Callers can be forgotten
A returning caller's profile is kept to serve their next call, expires after a set time, and can be deleted by number.
A person can step in
A staff member can watch a live call and take it over. The agent never chooses the number it transfers to.
What it does not do yet.
Gnani asks for honest limits, so they are on the front page and not in a footnote.
Measured on synthetic calls, not yet on recorded ones
Prisma and Timbre run live: the report card was measured on Gnani Prisma v3.0 on 6 Nov 2026. Its calls are spoken by Timbre, not by real callers, so it shows where the cliff is, not what a real caller will score.
Synthetic audio, simulated noise
The report card uses invented bookings spoken by synthetic voices, with noise and codecs added by software. Each condition is 24 clips, so one clip is about 4 points.
Digits are the hard part
A ten-digit phone number is the first field noise takes: heard whole in every clean, 8 kHz and codec clip, but in only 3 of the 6 clips that have one at 5 dB street noise. That is why the agent reads every number back and commits only on a spoken yes.
Missed-call callback not tried on a live number
The callback is built and tested with fakes: a Vobiz number whose Answer URL is a signed link is turned away unanswered and the front desk rings back. Vobiz itself, and a missed call on a real Vobiz number, have not been tried.
Calendar and inbound calls untested live
Google Calendar sync and calls received on your own number are written and tested against fakes, not against a real calendar or line. The bookings list itself is real: it reads the appointments store.
Caller Theatre ran on a scripted model
All nine scenarios pass with a scripted stand-in for the language model. That shows the tools, the schedule rules and the read-back work end to end; it says nothing about what the hosted language model does with the same calls, and its latency is not meaningful because the stand-in answers instantly.
Hours from words are rules, not magic
The plain-language hours reader handles English, Hindi and Marathi phrases it knows, shows what it understood for review and lists what it did not. A guess is flagged, and it never saves by itself.
Reasoning is not Gnani's
Replies come from a hosted language model. Document text and the caller's words reach that provider when it answers. When Gnani is down there is no second speech model: the caller hears a short held line and is offered a callback.
Not a legal product
Consent, labels and deletion are built to match the direction of India's rules on synthetic media and personal data. That is a design choice, not legal advice.
Try it
Make a call and correct a digit.
Open the desk, start a web call from your browser and give the agent a wrong number on purpose. It reads the booking back, you fix it, and you get a receipt anyone can check.
Team
Ayush Chougula
Builds the phone pipeline and every integration: Gnani, Vobiz, Google Calendar and messaging; schedules, booking and the noise report card
Atharva
Owns the UI: the desk, the phone layout and this site
Built for the Great Indian AI Internship Challenge by Gnani AI. Vakya is a student project and is not affiliated with Gnani AI. Every booking, caller and number on this site is invented.