First word in 41 milliseconds.

One stream, server-side: the first audio frame leaves the server 41 ms after new text arrives, and a minute of speech is made in eight seconds. Independent end-to-end measurements of rented voices land between 190 and 420 ms.

27 ms first audio
0.14 real-time factor
Server-side, no queue

Rented voices cost about 2.5 cents a minute. Ours: a tenth of a cent per 1,000 characters, falling to a hundredth.

A rented voice is rented compute plus a vendor margin: 2.5 cents (₹2) or more a minute. Ours is 10 paise, about a tenth of a US cent, per 1,000 characters today, about a minute of speech; half that in six months and under a paisa, about a hundredth of a cent, in about a year, because the model is ours and it runs on ordinary servers.

10pper 1,000 characters, today
5pin six months
<1pin about a year

Hindi, English (US), Hinglish. Then the world's languages.

Trained on how people actually talk: switching languages mid-sentence, numbers, amounts and dates read the local way, codes and acronyms spelled out. More languages are in training now.

हिंदी
English (US)
Hinglish
தமிழ்
తెలుగు
मराठी
বাংলা
ಕನ್ನಡ

No concurrency limit. Ever.

No queue and no scarce hardware: one four-core server makes 18 seconds of speech every second, and a server adds capacity in minutes, in our cloud or your data centre. The guard rails live inside the model: it speaks exactly the text it is given, reads names and amounts right, and stops cleanly at the end of the sentence.

No cap on streams
18× real time per 4 cores
Any server
Your DC or ours
Guard rails in the model

Give us the voice. We train the model for you.

Record your brand voice, an agent, an anchor, a founder, and we return a model trained on it: your timbre, your pronunciations, your pace. Same price, same speed, yours to run wherever you like.

30 min of recordings
Your model
Same price
Runs anywhere

Five convictions

41 ms first audio7× real time10p per 1,000 characters

Vartalaap · Pre-seed · Confidential

Pre-seed · October 2026 · Confidential

The voice stack you own. Live today, from $1.20 per million characters.

Our own model on our own engine: first audio in 27 ms in the engine (41 ms through the production API), no concurrency limit, about a tenth of the price of rented voices, on commodity CPUs, in our cloud, yours or your data centre. 22 voices live in English (US), Hindi and Hinglish; speech recognition and voice agents are next.

02 · The plan · Seven moves, one stack

We start with the voice. We end with the infrastructure.

Each move is the same model family on the same engine, a step further from the API and a step closer to the metal. Every step lowers the cost of the next.

01Text-to-speechLive22 voices · 41 ms
02Speech recognitionIn evaluationranked on real call audio
03Voice agentsBuilt · Englishanswers, acts, hands off
04Global languages3 languages liveIndia first, then the world
05On the deviceRuns on a laptop corephones, cars, kiosks
06Data centresPlan · month 30our own inference halls
07Sovereign AIBy designevery country's own stack

Vartalaap · Pre-seed · Confidential

03 · Why now · The problem, five times

The world's voice AI runs on rented models. We removed all five reasons.

Problem · the bill

Rented neural voices list at $15–80 per million characters, billed on every character before the call has said a word.

Problem · the language

Models trained on one kind of English stumble on code-mixed speech like “EMI ₹4,250 due है”, and on a language typed in another script.

Problem · the hardware

Every vendor queues on scarce accelerators, so capacity is rationed by a plan's concurrency limit instead of bought.

Problem · the data

A bank's or a hospital's customer voice leaves the building on every call, while data-residency law tightens from the EU to India.

Problem · the lock-in

Three vendors, three margins, three roadmaps. Nobody in the call path answers to the customer.

01/05

Ours · we own every model

$1.20 (₹100) per million characters today, half that in six months, under a tenth in a year.

10–70×below every rented neural voice’s list price

Ours · trained on how people talk

Code-mixed Hinglish read from Devanagari or roman letters; amounts, dates and OTPs read the local way.

2 scriptsone voice · v1.5

Ours · engineered for any compute

Our engine streams on commodity CPUs. No queue, no concurrency limit: a server adds capacity.

41 msfirst audio, new text, production

Ours · runs where the data lives

Our cloud, the customer's cloud or data centre, or the device. The weights travel; the voice stays.

0 bytesleave the building on-prem

Ours · one stack, one owner

Speech in, understanding, speech out: one company, one engine, one price that only goes one way.

1 vendorfor the whole call

Cost
Language
Compute
Data
Lock-in

Vartalaap · Pre-seed · Confidential

Chapter 01 · Text-to-speech · live in production

Voices that start speaking before the sentence ends.

POST api.vartalaap.si/v1/ttsStreaming

{
  "voice": "saanvi",
  "text":  ▍
  "speed": 1.0, "format": "pcm_24k"
}

0 ms to first audio

Real model output from production, unedited, 24 kHz. Tap a voice, then play.

0 msfirst audio, new textp50 · production API · p95 42 ms

0production voices18 English (US) · 4 Hinglish

7×faster than real timeone stream · RTF 0.14

$1.20per million characters(₹100) · 10 paise per 1,000 · today

Reads code-mixed text ₹4,250 · 12 अक्टूबर · lakh and crore · OTP digits one by one
Two scripts Hinglish typed in देवनागरी or in roman letters, one voice
Pace 0.5×–2× per request, pitch kept, streamed over REST or WebSocket
Your voice record it, we train it, same price, yours to keep

Vartalaap · Pre-seed · Confidential

Chapter 01 · The price · list prices, October 2026

Priced like infrastructure. 10 to 70 times below rented voices.

Text-to-speech · $ per million characters (₹) · ≈ 1,100 minutes of speechlinear · to scale

ElevenLabs Multilingual v2/v3$80 (₹6,720)67×

ElevenLabs Flash$40 (₹3,360)34×

Cartesia Sonic · Scale$37.4 (₹3,140)31×

Sarvam Bulbul v3$35.7 (₹3,000)30×

Google Chirp 3 HD$30 (₹2,520)25×

Deepgram Aura-2$30 (₹2,520)25×

Sarvam Bulbul v4 Flash$23.8 (₹2,000)20×

Amazon Polly Neural$16 (₹1,344)13×

OpenAI tts-1$15 (₹1,260)13×

Vartalaap today$1.20 (₹100)1×

Vendors' public pricing pages, 23 Oct 2026; ₹ at ₹84 / $. ElevenLabs $0.04 and $0.08 per 1k chars · Cartesia $299 per 8M credits · Deepgram, Google $30 per 1M · Sarvam ₹2 and ₹3 per 1k · AWS $16 · OpenAI $15 per 1M. Only legacy “Standard” voices (Google, AWS: $4) come within 4×.

Our price only goes one wayper million chars

Every cut comes from our own work, not a vendor's discount: a smaller model, a faster engine, a busier server.

Gross margin at $1.20Software

The model and the engine are ours, so the only cost under the price is an ordinary server's hour.

STT is already a commodity$0.003–0.016

per minute across Deepgram, AssemblyAI, Sarvam and AWS. The money in a call is in the voice and the agent.

Vartalaap · Pre-seed · Confidential

Traction · the first week of the portal

Three pilot customers. Twenty-one sign-ups in six days, without a launch.

Pilot customers

S

StayNowHotels · the voice concierge

Pilot

C

Converse AIPilot customer

Pilot

P

Platter HubPilot customer

Pilot

Sign-ups on tts.vartalaap.si · 16–21 Oct 2026no paid launch

0accounts

0outside companies

0days since the portal opened

GupshupHelo.aiTravon AIintegratedMeetwa AIRaaho gencodeDatavaveSmartDialProinfoDPD

Company names as entered at sign-up; Helo.ai, Travon AI, Meetwa AI, Raaho and gencode signed up from their company domains. Our own and test accounts are excluded. Travon AI has a live key and is calling the API.

Vartalaap · Pre-seed · Confidential

Chapter 02 · Speech recognition · next

Next: hearing people the way they actually speak.

Vartalaap STTwss · streaming · 8 kHz phone lineListening

Hinglishdigits kept as digitsamounts · dates · OTPspartial → final in one pass

The rule we already wrote down

Ranked on real human speech. Never on synthetic audio.

Our sibling QuickDial learned it the hard way: a benchmark voiced by TTS inverted the ranking of the models. Every Vartalaap STT decision is scored on human callers.

Where it stands

Built real-speech eval harness, Hindi + Indian English + Hinglish sets
Measured open baselines, including IndicConformer
Next fine-tune on call audio, digits/OTP/amount gate, then ship on the same engine as the voice

$23.1Bspeech recognition, 2030from $9.7B in 2025 · 19.1% CAGR · MarketsandMarkets

$104Bspeech & voice recognition, 203420.3% CAGR · Fortune Business Insights

Vartalaap · Pre-seed · Confidential

Chapter 03 · Voice agents · AgentBox

Then the whole call: an agent that verifies, acts and hands off.

Saral FinservCollections deskLive

नमस्ते राहुल जी, Saral Finserv से बात कर रही हूँ। आपकी EMI 12 अक्टूबर को due है, ₹4,250। Payment का plan बना लें?

Salary 15th ko aati hai. Can I pay on the 16th?

Bilkul. 16 October, ₹4,250, UPI link SMS pe bhej doongi. Confirm karne ke liye ek baar "haan" boliye.

Haan.

Account

Loan•••• 4471

EMI due12 Oct

DPD0 → 4

Amount₹4,250

Heard & filled

VerifiedDOB · last 4

ReasonSalary date

PTP

ConsentRecorded · 0:04

VerifyReasonPromiseLinkConfirmCRM

01

Answers from your scripts and products

Rate cards, policies, escalation rules. Nothing invented, nothing off-script.

02

Completes real actions in your systems

Promise-to-pay into the CRM, a card block in the core, a callback in the LOS.

03

Hands off to a person, with context

The transcript, the verified identity and the reason travel with the call.

04

Already built, already live in English

AgentBox answers hotel and restaurant calls for QuickDial in the US today: 253 ms from the caller's last word to the first audio, in the lab.

$3.5B → $35.2BAI voice agents, 2026 → 2033 · 39% CAGR · Grand View Research

$7.16 vs $0.07–0.15a human-handled inbound call vs a voice-AI minute · MarketsandMarkets, 2026

80%of common service issues resolved by agentic AI without a human by 2029 · Gartner, Mar 2025

Vartalaap · Pre-seed · Confidential

Chapter 04 · Languages · India is the beachhead, the market is global

Three languages live. Then every market rented voices serve worst.

01 · Live today

English (US), Hindi, Hinglish.

18 US-English voices and 4 Hinglish voices in production. Hinglish is the hard case, two languages in one sentence typed in either script, and the proof that our recipe travels to a new language.

22production voices528MHindi mother-tongue speakers · Census 2011

02 · Beachhead · India

First market, first proof.

India is where we sell first: three pilots, 21 portal sign-ups in a week and a price local buyers can carry. Bengali, Marathi, Telugu, Tamil and Gujarati are the next five by speakers.

958Minternet users · IAMAI 20251.09Bbroadband connections · TRAI

03 · Then · the world

The markets rented voices serve worst.

The same engine and the same price, where rented voices are thin and dear: Arabic in the Gulf, Bahasa in Indonesia, Spanish and Portuguese in the Americas, Swahili in East Africa, Turkish.

7,000+living languages · Ethnologue252Mspeak Indonesian

livebeachhead · nextthen

Live today
Beachhead
The world

Vartalaap · Pre-seed · Confidential

Chapter 05 · Edge · the same voice, on the device

Small enough for a phone. So the voice never waits for a network.

one model

Phones486M GenAI phones in 2026

Carsin-car assistants $8.4B

Kiosks & countersbanks, rail, hospitals

Wearablesglasses · 14.3M in 2025

TVs & set-top boxesthe living-room voice

Feature phones & IVRevery phone line

Why we can

A 6-layer model distilled from a 24-layer teacher.

Every Vartalaap voice is a small student model trained to sound like a large one. It already runs faster than real time on a single laptop core: a minute of speech in under three seconds of one core.

0×real time on one laptop coremeasured · Apple M-series

0 msnetwork on the deviceno round trip, works offline

$118.7Bedge AI, 203321.7% CAGR · Grand View Research

45%of 2026 phones run GenAICounterpoint, Jun 2026

Vartalaap · Pre-seed · Confidential

Chapter 06 · Data centres · our own inference halls

At scale, the cheapest minute is the one we run in our own building.

vtts-hall-1your region · inferenceServing

0servers

0live voices at once

$1.20per million chars

What we know today, measured

One ordinary server speaks for 14 to 18 callers at once.

Measured on a 4-core cloud server: 14 seconds of speech every second through our production API, 18 in the engine's own bench, 59 ms to first audio with four people talking. No accelerator in the path, so capacity is bought by the rack, not rationed by a queue.

Why our own halls

Inference is the load about two-thirds of AI compute in 2026, up from a third in 2023 · Deloitte
CPUs are bought, not allocated no accelerator queue, so a hall is sized by demand in any region
The first market is building India alone: 2.2 GW → 12 GW of data centres by 2030 · Wood Mackenzie

Vartalaap · Pre-seed · Confidential

Chapter 07 · Sovereign AI · and everything after

Every country wants its own voice stack. We can hand it the weights.

01

Weights you can hold

The model ships to your hardware: a bank's data centre, a ministry's cloud, an air-gapped hall.

02

No accelerator to import

It runs on the servers a country already has, so sovereignty doesn't wait for a GPU allocation.

03

Built for data-residency law

From the EU's GDPR to India's DPDP Act (in full by May 2027) and RBI's payment-data rule: the model runs where the data must stay.

04

Every citizen, every language

Helplines, farm advisories, hospitals, courts and counters: public services that answer in the caller's own language.

>$30BNVIDIA sovereign-AI revenue, FY26, 3× a year earlier

$1.2B(₹10,372 Cr) IndiaAI Mission, five years · MeitY

€200BEU InvestAI, incl. €20B for AI gigafactories

Vartalaap · Pre-seed · Confidential

Market · global, with India as the beachhead

A $16B voice market in 2025. About $52B by 2030.

$52BGlobal TAM · 2030

$29BSAM · voice out · 2030

$2.3BIndia beachhead SAM

$23MSOM · ARR · M36

01 · Top-down · the layers we sell

Three published markets, one stack.

Text-to-speech · AI voice generation · 30.7% CAGR · MarketsandMarkets	$4.2B → $15.8B
Speech recognition · 19.1% CAGR · MarketsandMarkets	$9.7B → $23.1B
AI voice agents · 39% CAGR · Grand View Research	$2.5B → $13.1B
= TAM, 2025 → 2030 · ~26% a year	$16.4B → $52B
SAM: what speaks · TTS + voice agents, 2030	$29B

2030 for TTS and voice agents is our arithmetic from each publisher's own CAGR (they publish 2031 and 2033). Wider definitions run higher: conversational AI $17.1B (2025) → $49.8B (2031), MarketsandMarkets; $41.4B by 2030, Grand View Research.

02 · Bottom-up · our SOM

Sized from customers, not from share of a forecast.

300 enterprise and platform customers × 3M agent-minutes a year × $0.02	$18.0M
TTS API: 20B minutes of speech (18T characters) at <$0.12 per million	$2.1M
On-prem and sovereign licences: 5 × $0.5M	$2.5M
= ARR at month 36 (≈ ₹190 Cr; the roadmap plans ₹192 Cr)	$23M

Plan assumptions, not measurements. $23M is 0.2% of the 2030 voice-agent market; 900M agent-minutes is 0.07% of the ~1.4 trillion minutes the world's 15.3M contact-centre agents talk a year (Cavell 2025 × 90,000). India beachhead: 1.3M agents (NASSCOM, unverified) × 90,000 min × $0.05 = $5.9B; 40% in our languages and sectors = $2.3B.

03 · The category is proven

Voice is where venture money already went.

$11BElevenLabs · Feb 2026 · $500M+ ARR

$15.8BSierra · May 2026

$4.5BDecagon · Jan 2026

$1.5BSarvam · Jun 2026 · India

$1.3BDeepgram · Jan 2026

$750MPolyAI · Dec 2025

Voice-AI equity: $2.1B in all of 2024, more than $7B in Q1 2026 alone.

Top-down
Bottom-up
Comparables

Vartalaap · Pre-seed · Confidential

Roadmap · 36 months · the plan

Every six months, a new layer. At thirty-six, our own halls and sovereign deployments.

NowText-to-speech live22 voices · Hinglish v1.5
M3Speech recognitionHindi · Hinglish · English
M6AgentBox in Hindifirst enterprise pilots · ARR $0.6M (₹4.8 Cr)
M12+5 Indian languagesARR $2.3M (₹19 Cr)
M18Global languagesArabic · Bahasa · Spanish · ARR $5.7M (₹48 Cr)
M24On-device SDKphones · cars · kiosks · ARR $11.4M (₹96 Cr)
M30First own hallfirst region · inference
M36Sovereign deploymentsARR $23M+ (₹192 Cr)

A plan, not a measurement. ARR = MRR × 12 on the curve the v1 deck set out in rupees (₹40 L → ₹16 Cr MRR); dollars at ₹84 / $. Each step reuses the model family and engine already in production.

Vartalaap · Pre-seed · Confidential

Competition · who owns the voice

Others rent the world a voice. We sell the stack you own.

Company	TTS · $ / 1M chars	First audio · claimed	Valuation
ElevenLabsUS · UK	40 – 80	~75 ms · Flash	$11B
CartesiaUS	37	<90 ms	n/d
DeepgramUS · Aura	15 – 30	<200 ms	$1.3B
OpenAIUS · tts-1	15 – 30	—	—
Google CloudChirp 3 HD	30 (legacy 4)	—	—
Amazon PollyAWS	16 – 30 (std 4)	—	—
Azure SpeechMicrosoft	~15 – 16 *	—	—
SarvamIndia	24 – 36	—	$1.5B
Vartalaappre-seed	1.20 → <0.12	27 ms engine · 41 ms API	pre-seed

yespartlybuildingnoPublic pricing and product pages, 23 Oct 2026; latency as each vendor states it. * Azure: third-party figure, unverified. PlayHT: team acquired by Meta, Jul 2025; API closed. Valuations: latest reported round.

Price

10–70× below every rented neural voice

Because the model, the engine and soon the building are ours.

Compute

Commodity CPUs, then devices

No accelerator queue and no concurrency limit between a customer and capacity.

Ownership

Weights that can leave with the customer

On-prem and sovereign from day one, not an enterprise add-on.

Vartalaap · Pre-seed · Confidential

The raise · pre-seed

Raising our pre-seed. The terms open in the room.

Target raise

Instrument

Runway

Use of funds

Milestones it buys

Next round

Sealed · shared in the room

The raise, the terms and the use of funds open when we talk.

30-minute intro call

Hear it live, then talk numbers.

With Rajat Mangotra, CEO, and Puru Chaudhary, CTO. Pick a slot and the invite lands in your calendar with a video link.

The voice, live, in English, Hindi and Hinglish
Unit economics, the stack and the 36-month plan
The terms, opened

Book 30 minutes Prefer email? roshan.purush@gmail.com

Vartalaap · Pre-seed · Confidential

Founders

Two founders, one stack. One runs the company; the other trained every model in it.

Rajat MangotraFounder & CEO

Runs Vartalaap and scales it, from the company and its customers to the pilots and the money. He has spent eight years building AI and ML products across retail, e-commerce, fintech, education and hospitality, with a record of making systems faster, cheaper and bigger, often cutting compute costs by 90% or more. Most recently, as co-founder of QuickDial AI, Vartalaap’s US sibling, he helped bring a CPU-only voice-agent stack down to a quarter of a cent a minute. An alumnus of IIT Kanpur.

Purushottam (Puru) ChaudharyFounder & CTO · AI inference engineer

Built the voices, the engine they run on and the products above them. Fifteen years of software for GE HealthCare, S&P Global, Bristol Myers Squibb and State Street; three years training models; three companies founded. Author of The Inference Mechanic.

LinkedIn The book

22voices in productiontrained in-house

4Hinglish releases in a weekv1.1 → v1.5, each through a gate

0.25¢the minute QuickDial reachedUS sibling

1,400pages on inferenceThe Inference Mechanic

Vartalaap · Pre-seed · Confidential

A message from the founder

Puru ChaudharyFounder & CTO

Puru ChaudharyFounder & CTO, Vartalaap

Vartalaap · Pre-seed · Confidential