# Platform

Describe the call. Hear the agent. Then let it call.

Type the job. Review the diff. Talk to it in the browser. Then publish a version that already knows when it is allowed to dial.

PLATFORM

Build

Describe the call. Accept the agent.

TEACHER AI

Describe · Clarify · Configure · Accept

Checkpoint 1

✓

Automatic Context Pull

Describe your agent in plain English…

Type your requirements…

→

Prompt DiffRecommended Settings

Diff appears as Teacher AI drafts…

CancelAccept & Save

Teacher AI

Describe the call. Review the diff. Accept the agent. The teaching chat is typed. Speaking happens afterward, in the live test.

A sentence, or a starter such as lead qualification, appointment booking, customer support, or product consultation
No prompt yet: Vaani drafts one. A prompt already exists: the same chat edits it and changes as little as possible
Clarifying questions first, then a sentence diff — accept only the sections you want

VOICE AGENT

Listen · Reason · Speak · Act

Idle

Ears

STT

Brain

LLM

Mouth

TTS

Actions

Tools

Memory

RAG

Agents

The studio around that chat is the rest of the agent: who it is, what it can do, and how you try it before anyone is dialed.

Persona, senses, cognition, tools, experience, and analysis
In-browser chat and audio test, with a transcript
Create from scratch, from a template, from a clone, or from an exported config

TEMPLATES

Browse · Preview · Adopt · Customize

Gallery

Healthcare · Clinic Reception

Inbound · English · Booking, triage, warm Hindi/English.

Female

Real Estate · Lead Qualification

Outbound · English · Qualify buyers, book property visits.

Female

E-com · Order Support

Inbound · English · Verify orders, delivery updates.

Male

Templates

Start from a job that already works. A template is the whole call — not a prompt snippet.

Prompt, guardrails, voice, hold and voicemail behavior, tools, and the fields to extract after the call
Filter by inbound or outbound, and by industry
Keep a private template for your organization, build one in a five-step wizard, or clone a live agent

VOICE LAB

Browse · Audition · Clone · Assign

Explore+ Clone

LanguageAccentProviderGender9,400+ voices

Aaden

Friendly American middle-aged male.

▶

MaleEnglish

Aarti

Warm low-latency voice for live calls.

▶

FemaleHindi +1

Palak

Clear, consultative sales tone.

▶

FemaleEnglish

Meera

Native regional dialect specialist.

▶

FemaleTamil

Rohan

Calm collections specialist, firm but fair.

▶

MaleEnglish

Sofia

Bright hospitality concierge voice.

▶

FemaleSpanish

Clone Your Voice

Upload 15–30s of clean speech. Clones stay private to your organization.

♫

brand-ambassador.m4a

22s · ready to clone

Clone ›

Voice Lab

Choose how it sounds, how fast it talks, and what it does when the line goes quiet.

Audition a voice, clone one from an upload or a recording, or register a voice you already own
Background noise, filler words, a breathing sound, and eagerness: slow, balanced, or fast
After silence, a pulse check — then hang up if the caller stays idle

KNOWLEDGE BASE

Ingest · Clean · Embed · Retrieve

FastDeep FAQ

Website

PDF + OCR

Word

Paste

Clean

Normalize

Embed

Vectors

Retrieve

On-call

ON-CALL RETRIEVE

—

Caller · What's your return window on unused items?

CITED · return-policy.pdf

Unused items can be returned within 30 days of delivery with original packaging.

Paste
pricing-faq.md
Queued
PDF
return-policy.pdf
Queued
Web
docs.company.com
Queued

Compiling index

Knowledge Base

Your docs, policies, and site become answers on the call, shared across agents.

Text, PDF, Word, and a URL. Text is stored immediately; files and pages are read in the background
Pick fast or slow when you create the base. That choice stays
Attach one base to more than one agent. Detaching it does not throw the answers away

PLATFORM

Humanise

How it sounds when the line is live.

FILLER WORDS

Acknowledge · Mask · Continue

Caller is speaking…

CALLER

Reschedule my dental visit to Tuesday…

···

AGENT

Checking the schedule for you…

0:00

Lookup masked

Got itUh-huhHaanjiSure

Filler Words

Active-listening cues and conversational affirmations—Got it, Uh-huh, Haanji—so agents nod in real time, mask tool latency, and never leave callers in dead air.

Backchanneling that signals the agent is hearing you, not hanging
Sub-second latency masking before CRM, calendar, or KB lookups
Master toggle, 0–1 frequency, and custom multilingual filler tags

BACKGROUND AMBIENCE

Layer · Mix · Stay Clear

Office under the voice — outbound only, so speech-to-text stays clean.

Office
Support floor
Quiet room
No Noise

VOICE

Office

0.10

Sweet spot
0.05–0.20

Background Ambience

Background noise and a breathing sound sit under the voice, so the line does not sound like a silent booth.

Office or room tone, plus a breathing sound when you want the agent to feel present
Turn it off when the call should be quiet
The mix stays under the voice — it does not replace what the agent says

EAGERNESS TO SPEAK

Wait · Detect · Take the Floor

Caller is speaking — the gauge fills only after silence hits the threshold.

0ms

Balanced

FastBalancedSlow

My account number is 4-5-8…

Eagerness to Speak

Choose how fast it talks: slow, balanced, or fast. Idle handling is separate, and concrete.

Slow, balanced, or fast — how quickly the agent takes the turn
After silence, a pulse check. If the caller stays idle, the agent hangs up
Tone is not its own slider. It comes from the prompt plus these controls

EMOTION & MOOD

Sense · Soften · Resolve

Caller sounds furious, but the agent replies with the wrong tone.

Caller · You charged me twice — fix this now!

“Hope you’re having a wonderful sunny day!”

Wrong tone for this caller

CALL ARC

Negative

Emotion & Mood

Emotion follows the conversation. It is not a tone slider sitting next to the voice.

Tags shift with the caller, from the prompt and the turn — not from a single mood dial
The voice you assigned in Voice Lab is what they hear
Eagerness and idle handling stay their own controls

NATURAL BARGE-IN

Hear · Cut · Pivot

Both sides stream at once — the agent can hear while speaking.

AGENT · OUT

Our office hours are Monday through Friday from nine…

FULL DUPLEX

CALLER · IN

Idle

…

Waiting for a real interrupt…

Natural Barge-in

Full-duplex interruption handling—callers can cut in mid-sentence, the agent flushes audio in under 150ms, and pivots without steamrolling or clunky apologies.

Sub-150ms playback flush with LLM cognitive pivot on new intent
Filters coughs, noise, and passive uh-huhs from real “wait / stop” interrupts
Interruptible greetings when callers already know what they need—or protected ones for disclosures

NATURAL CODE-MIX

Detect · Blend · Stay Native

Caller mixes Hindi and English in one sentence — no language menu.

EN termRegionalBlend

CALLER

Merapaymentcompleteho gaya,lekinconfirmation mailnahi aaya.

AGENT

…

KEEP IN ENGLISH

Video KYCInvoiceCredit LimitService Slot

Natural Code-mix

Speak and understand the way people actually talk—Hinglish, Tenglish, and other blends that switch mid-sentence, keep English jargon intact, and never force Press 1 for English.

Mid-turn switches with colloquial grammar—not stiff textbook translations
STT auto-detect plus Voice Lab TTS that pronounce both languages natively
Bilingual fillers and emotion stay coherent across language shifts

PLATFORM

Connect

Act on the call, or start it from code.

INTEGRATIONS

Bridge · Act · Sync

Cal.com

Sheets

VAANI

Transfer

—

Integrations

The agent can act on the call: book a time, message the caller, write the outcome, or hand the line to a person.

Cal.com availability and booking
WhatsApp follow-ups, and Google Sheets for the outcome of the call
Schedule a callback, end the call, transfer to a person, or transfer to another agent
The knowledge base you attached is what the agent retrieves while it talks

TOOLS

Intent · Execute · Speak

CALLER

Can you check delivery on order #9821?

check_order_statusGET · /orders/9821

“Got it, let me check that…”

···

status Out for deliverycourier BlueDarteta 2:30 PM

AGENT

Out for delivery with BlueDart — around 2:30 PM.

transfer_callbook_appointmentsend_whatsapphold_call

Tools

Built-in actions cover the common jobs. Anything else is an HTTP function you register with a name, a description, and an endpoint.

Cal.com, WhatsApp, Sheets, callback, end call, and both kinds of transfer
Custom tools run mid-call and the result is spoken back
Your own key for speech, the language model, and voice — yours only, yours with a platform fallback, or the platform key
A custom language model on your websocket is not billed as platform model usage

WEBHOOKS

Stream · Sign · Deliver

POST call_started

SIP dial

POST call_ringing

180 Ringing

POST user_picked_up_at

Audio open

POST call_ended

42s · BYE

POST call_postprocessing

Transcript · entities

YOUR BACKEND

/webhooks/vaani

Waiting for events…

X-Webhook-Signature

Webhooks

Each agent or campaign can stream the life of the call to your endpoint, with the events you choose and a secret you verify.

Per agent or per campaign — pick the event list
Signed so you can tell a real Vaani event from anything else
Includes the post-call payload: transcript, disposition, and extracted fields

API

One agent · Any channel

YOUR APP

Backend

X-API-Key

VAANI

Phone

PSTN / SIP

In-app

WebRTC

Ready

API

A developer can start the same agent without the studio. Trigger a call with an agent id, a number, and metadata.

Phone call or in-browser session, against api.vaanivoice.ai
Pass variables in metadata, and override the voice or caller ID for that call
Skip the DND check on purpose, or run in test mode without dialing
Speech, reasoning, and voice each have a primary and a fallback

API Reference

PLATFORM

Run

Dial only when the call is allowed.

TELEPHONY

Managed · BYOC · SIP

LINE

+91 98765 43210

Idle

Vaani ManagedWallet · Instant DID

Inbound

→ Support agent

Outbound

→ Campaign dial

CALL PROGRESS

Dialing

Trying

Ringing

Answered

Concurrent calls

0 / 50

Warm transfer

SIP REFER · Bridge

Stereo record

Caller · Agent

TRAI hoursDNDTLS SIPUDP / TCP

Telephony

Outbound calls and inbound numbers use the same agent. Bring a number you already have, or take one from a carrier Vaani already connects.

Numbers from Vobiz, Twilio, or Plivo
Or bring your own SIP number over UDP, TCP, or TLS
One agent on the way in and the way out

CAMPAIGNS

Upload · Dial · Convert

UploadScrubDialRetrySync

FIRST TURN

Namaste Priya, aapka EMI ₹12,400 due 25 Sep…

Q3 Renewal Reminders

Preparing

Leads

10k

Done

—

Connect

—

Progress0 / 10,000

10:00–18:0050 parallelRetry ×3 · 30m

41%

Confirmed

13%

Callback

22%

Busy

24%

No answer

Campaigns

Upload the list. Vaani dials only inside the window, skips DND, and retries the rest.

A CSV or a Google Sheet, with a pace you set and retries for contacts that fail
Pauses outside the daily window and the days you selected, then resumes when the window opens
India calling stays inside 08:00–20:00 IST, including on a single call from the API
DND is checked before dial. Test mode places the call without dialing

ORCHESTRATION

Gate · Loop · Fan-out

PythonC++

Trigger

Wallet · Agent · Slots · DND

Live loop

VAD → ASR → LLM → TTS

Fleet

Windows · Pacing · Retry

Fan-out

Audio · Eval · Webhook

Barge-in <150msRedis slotsStuck recovery

Orchestration

The same agent runs the live conversation: listen, answer from what you attached, act, then speak — while the campaign stays inside its rules.

DND and the calling window are checked before the line opens
Personalization from the list shows up in the first turn
Lifecycle email when the campaign moves, and a write-back of what the call extracted

RELIABILITY

Failover · Watch · Heal

Primary

SIP trunk A

Standby

→

Fallback

SIP trunk B

Standby

AGENT WORKER

Primary container

Healthy

WATCHDOGS

Idle air

Reprompt

Stuck LLM

Filler mask

Duration

Hard cap

Standing by

15-min sweeperRedis rollbackDurable audio

Reliability

A call that sticks, or a carrier that fails, should not stall the list behind it.

Calls stuck in progress for more than 15 minutes are recovered
Speech, reasoning, and voice can fail over so one provider does not end the call
Warm and cold transfer show up in call history, including the transfer leg

PLATFORM

Improve

Hear it first. Then read the call.

TESTING

Web · Phone · Chat

Web

WebRTC · $0

Phone

Live PSTN

Chat

Silent debug

BROWSER MIC

Ready

Listening · Thinking · Speaking

MOCK VARIABLES

{{name}} Rahul{{balance}} ₹12,450{{due}} Tomorrow

Webhook simVoice LabKB retrieve

Browser test

Talk to the agent in the browser, by chat or audio, with a live transcript, before you publish.

Chat or audio, in the studio, before the agent is on a real number
The transcript stays with the session so you can see what it heard
The API can run the same agent in test mode without dialing

ANALYTICS

Contain · Convert · Observe

81%

Containment

68%

Connect

48s

AHT

DURATION MIX

<10s

10–30s

30s–1m

1–3m

>3m

78%

Appointments

62%

Payments

84%

Qualified

STT 118msTTFT 240msTTS 165ms

Turn 523ms

After the call

Read what the agent extracted from real calls — the disposition, and the fields the template defined.

Disposition is scored after the conversation, not guessed on a slide
Extracted fields follow the template: the questions you already decided to keep
Transcripts stay attached so you can check the score against what was said

PLATFORM

Scale

Publish a version. Keep the campaign moving.

CONCURRENCY

Reserve · Cap · Pace

ACTIVE SLOTS

0 / 50

Standing by

Hard cap 50Grace 55

Inbound

Outbound

Transfer

API / Inbound

Priority queue

Campaigns

Throttle dialer

Redis LuaTTL cleanupParent · Child

Concurrency

Publish a version. If a call worker dies, the line frees itself and the campaign keeps going.

Each live call holds a slot and renews it while the agent is up
If the worker dies, the slot frees itself and the dialer continues
Each campaign keeps its own pace, separate from the rest of the account

FALLBACKS

Fail · Step · Continue

Model

Primary → Fallback AI

Trunk

Carrier A → Carrier B

CLI

5-level number cascade

Worker

Standby HA pod

Guard

Filler · Reprompt · Transfer

Curfew

Queue → 08:00 dial

Docs

Text → OCR

Standing by

Fallbacks

Speech, reasoning, and voice each have a fallback, so one provider failing does not end the call.

A primary and a fallback for speech-to-text, the language model, and the voice
Your own key can stand alone, fall back to the platform, or stay on the platform key
Warm and cold transfer remain visible in the call, including the transfer leg

VERSIONS

Tinker · Switch · Roll back

Editing draft…

AGENT

Support Concierge

Live build: v3 · Current

Live

Prompt · Voice · Tools

Softer greeting
New transfer rule

Last stable publish

Original opener
Strict transfer

WHAT CHANGED

+ Greeting warmed for Hinglish

~ Transfer threshold lowered

Keep currentSwitch to previous

Versions

The live agent does not change on every keystroke. You publish a version, and you can go back.

Published versus edited, so a draft stays a draft
Revert to an earlier version after a prompt, voice, or tool change
Export and import the config as JSON

Hear a Live Call Talk to sales