maya research
The opportunity
The fundamental shift
Voice is becoming the universal operating system for everything.
One way to run every device, machine, and workplace — just by talking to it.
01
Personal Devices
- Smartphones
- Laptops
- Smart glasses
- Earbuds
- Wearables
02
Humanoids & Robots
- Humanoid robots
- Home robots
- Service & delivery bots
- Companions
03
Home AI
- Home assistants
- Smart appliances
- Ambient home OS
04
Enterprise Workflows
- Call centers
- Field agents
- Retail staff
- Sales teams
- Support teams
The future
For 5 billion people restricted by traditional text interfaces.
Voice is the inevitable interface and Maya is capturing their gateway to the digital world.
They talk to a phone the way they talk to a person, in their own language.
Payments
Finance
Shopping
Learning
Creation
1B
served by today's internet — English, typed
5B
the next five billion — they talk, not type
maya research
The scaling factor
The next 5 billion people think and speak in Thousands of native languages.
The multi-trillion-dollar shift to voice stays dead until the AI converses exactly like a native.
How today's AI sounds
Dead on arrival.
The moment it sounds foreign and doesn't think like them, no one can talk to it.
0:00
The moment it speaks like them
The universal voice interface.
It becomes a universal voice interface, the way everyone runs technology.
We just needed $1M to train the most native model, out-engineering labs with 100x the resources.
Source: Voice Arena ↗
maya research
Hear it for yourself
Hear it for yourself
Native vs Alien
Listen to Maya 2: Our most native voice model yet
Our model
Maya 2 Native
0:00
Cartesia Sonic
0:00
Sarvam Saaras v3
0:00
A fraction of the big labs' budgets, and already outpacing them on quality.
maya research
The technology
The Architecture
We applied the DeepSeek playbook to native voice.
The whole stack is engineered to be cheap to train and cheap to serve. Three ideas do most of the work.
01 · Architecture
HARP
Hybrid attention + Parallel decoding. One checkpoint: block-diffusion + self-speculative, with SSM routing that keeps long turns linear-cost. Quantized, single GPU.
02 · Backbone + data
Continual Pre-Training
We take full leverage of frontier text models — A Strong Text based backbone that already reasons across languages — and continually pre-train it on trillions tokens of audio we own at source.
03 · The data flywheel
Data Advantage
Our consumer app is a data engine — every conversation feeds proprietary, accented speech back into training. Large teacher models add unlimited synthetic data. No one else has this loop.
The result: The most native speech model, which is at frontier quality and a fraction of the cost
The flywheel
A continuous data flywheel of native conversations from our app is training our models to own the interface.
We have already crossed 34 million conversations. Every new interaction generates proprietary data that makes our models sound native.
34M
real conversations, compounding every day, in Hindi, Telugu, Bengali and 20+ more languages.
↳ tap to see the data
01
People talk
02
Native data
03
Better models + memory
04
More users
maya research
Business model
Enterprises already build on our models.
Our latest models will be the biggest driver of revenue — the same models enterprises already run, opened up as a paid API.
Half a million downloads
Developers already chose us
Maya's open-weight models, downloaded by developers on Hugging Face ↗.
The biggest inference providers host our models
Served where enterprises build
Already live commercially — FAL ↗.
Only company from India
On the leaderboard
The only Indian company on the Speech Arena leaderboard — Speech Arena ↗.
How the models earn
Voice agentsthe platforms building on speech
Enterprisesdirect, usage-based API
Inference providershosted where they already build
The engine · Maya S2S
A custom speech-native architecture, built for the next five billion.
Built end to end for the world's dialects, emotion, and long conversations — watch one moment travel through it.
01 · Perceive
Hears speech as speechnever text
● Meena speakinglistening
3 parallel encodersacoustic · semantic · prosody
intonation
dialect marker ✓
emotion: उत्साह (excitement)
hesitation ✓
Semantic tokenizerlow-rate · align
◍◍◍◍◍◍◍◍
Duplex state estimatornot VAD
user continuing0.00
user done0.00
self-echo0.00
barge-in0.00
02 · Understand
MoE reasoning coreroutes by function, not language
Fused speech state — what she meansunderstood, never transcribed
“मुझे शादी के लिए हरी सिल्क साड़ी चाहिए, हल्की, ₹3,000 के अंदर, और शनिवार से पहले।”
Memory · speech embeddingsloaded every turn
how she talks
past corrections
her preferences
long context
Router · input-type tagsemotional + factual
Dialect & accent
her Hindi, her way
Emotion
wedding joy
Memory recall
it already knows her
Reasoning
≤₹3,000 · by Saturday
Turn-taking
she has yielded
+ 12 experts
routed by function
▶ Speech-act frame — the plan, not the sound
intent
match + confirm
prosody
bright · warm
confidence
0.94
interruptible
yes
03 · Respond
Streaming generator3 lanes at once
Audible lanebecomes sound
◍◍◍◍◍◍◍
Silent reasoning lanethinks while speaking · never spoken
◍◍◍◍
Control laneprosody · pacing
◍◍◍
i-MiMi codec · multi-codebook48 kHz
CB1
CB2
CB3
CB4
+
⚡ speculative head — first packet fast
Rendering layervoice identity · decoupled
Fast causal vocoder
first packet · barge-in
Flow-matching vocoder
longer tail · quality
● Maya speakingcomposing…
↻ Maya hears her own voice — echo vs. barge-in
నువ్వు రారాthe same three words are a warm invite, or a challenge to fight·Maya hears which
Understands meaninghesitation, sarcasm, urgency, accent, and tone
Natural interruptionsinterrupt, laugh, correct, or talk over it, and the talk flows on
Speaks like a localright rhythm, stress, dialect, emotion, not just pronunciation
Thinks while speakingreasons, remembers, and prepares the next step mid-sentence
Handles long contextthe whole conversation stays live, however long it runs
maya research
The landscape
Vertical integration
We are completely focused on one thing: Monopolizing the interface between the next five billion people and technology
Blue ocean
Vertically integrated, model + app ↑
↓ Model / API layer only
◂ Global · English-first
Focus on the rest of the world ▸
OpenAI
Sesame
Maya
ElevenLabs · Cartesia · Moshi
Smallest AI
Sarvam
Owning both app and model layers enables us to iterate very fast and capture more value.
maya research
The ask
The opportunity
We have raised $1.9M so far. Now $8M can make Maya the default.
$1M
Spent — and we already built
- Models that beat OpenAI's latest at Elo
- The most native speech model, 31 languages
- The most accurate understanding model
- Only company from India on the leaderboard
- 3M app downloads · 500K model downloads
- Inference platforms hosting us
→
$8M
What $8M unlocks
- Ship Maya 3, a speech-to-speech (S2S) model: it listens and talks in one model, so conversations feel real, not robotic turn-taking
- Improve app & grow
- Kickstart enterprise motion & monetization efforts
To achieve the above, we'll invest heavily in a strong tech, product & sales team.
Maya is becoming the default voice interface for the next 5 billion.
1 / 21
The data
What's actually inside 34 million conversations.
How we collect it
Every conversation inside the Maya app — consumer and enterprise — opted-in, real usage, not scripted studio recordings. No paid data-labeling vendors.
What kind of data
Full-duplex audio, transcripts, prosody, accent, and code-switching context — across Hindi, Telugu, Bengali and 20+ more languages.
Quality
Real accented speech from real people, not synthetic TTS loops. Filtered and deduplicated — the exact distribution our models are judged on.
How it helps the models
Feeds continual pre-training and the data flywheel directly — every new conversation makes the next model more native.
Other nuances
Large teacher models add unlimited synthetic augmentation on top of this — but the real accented core is what no one else has.
Organic content on YT
People are already using Maya on their own and making how-to videos about it — see the organic content on YouTube ↗.
The next five billion
Maya is becoming the default voice interface for the next 5 billion.
A world of languages
The next five billion speak 7,159 living languages.
A glimpse of the world's living languages · green = the handful today's AI serves · the rest are the next five billion.
In India alone: 1,369 mother tongues, 121 major languages — 22 with any institutional support.
Quality per dollar
Best model anyone can compete with — on 1/100th the capital.
VoiceArena Hindi score
Leaner · less capital raised →
1086
1072
1058
1044
$10B
$1B
$100M
$10M
The cracked corner
Maya$1.9M
Smallest$8M
ElevenLabs$781M
Cartesia$191M
Sarvam$275M
MiniMax≈$1.77B
xAI≈$42B+
Maya delivers frontier voice quality on a fraction of the capital. Every rival raised more—some by four orders of magnitude.
Quality: VoiceArena Hindi (Bradley-Terry Elo). Total disclosed capital through 4 Jul 2026: Maya $1.9M · Smallest $8M · Cartesia $191M · Sarvam $275M · ElevenLabs $781M · MiniMax ≈$1.77B incl. IPO · xAI ≈$42B+ incl. debt.
Why it doesn't split focus
Enterprise is the same model — turned on, not built.
No new build
Same weights, same training loop that already powers the consumer app. Enterprise is a revenue tap, not a second product — zero diverted engineering.
Demand already exists
500K downloads of the open weights — inbound from conglomerates and startups, already hosted on AWS, Cloudflare, FAL and Baseten. Nothing to cold-start.
Why sequence it
We turn on the paid API once the model matures — so it funds the consumer mission instead of forcing another dilutive raise.
Focus stays consumer
One team, one asset, one focus. Enterprise revenue rides on top of the flywheel — it strengthens the consumer bet, doesn't compete with it.
Open-weight leaderboard
Maya‑1, our open-weight model — released 13 months ago, still top 5.
Source: Artificial Analysis ↗