v4.2 · streaming · 142 ms median

A voice that knowshow to listen.كيف يُصغي.如何倾听。कैसे सुनना है।comment écouter.jinsi ya kusikia.cómo escuchar.how to listen.

Nuri is a voice intelligence platform built for the moments that matter — a whispered command, a child's question, a robot crossing a noisy floor. Real-time, on-device, fluent across dialects most models forget.

Trusted in production by

Try it · no login, no install

Speak. Watch Nuri think.

Tap to speak · simulated

nuri-live · streaming— ms
Press the orb to start a simulated session.
000 ms
median round-trip
000
languages
0
data leaves device

Manifesto · 01

The next interface is not a screen. It is a conversation that remembers you.

— Nuri research, Cairo · Singapore · SF

We don't build voices to perform. We build them to listen — first to silence, then to a breath, then to the dialect spoken in the kitchen where no white paper has ever been written.

Every interaction stays on the device. Every memory survives the night. Every latency budget we missed got rebuilt until 142 milliseconds felt like a held breath, not a queue.

This is voice as quiet infrastructure — for healthcare, for robotics, for the games your kids play, for the elder who deserves to be understood the first time.

Latency you can feel.

A friend who already knew the answer, and waited for you to ask. Local memory that survives the night. Turn-taking under the threshold of a held breath.

142 msmedian round-trip · on-device
nuri · live9:41
Hey, can you remind me to call Maya this evening?you · 0.0 s
Set for 6:30. Want me to pull her last note before you call?nuri · 142 ms
Yes. And read me the headline.you · barge-in
“Khaleeji rollout: nine days, 97 % satisfaction.”nuri · 138 ms

On-device pipeline

01ASR
streaming · 16 kHz · WebRTC
38 ms
02Memory recall
on-device · sub-graph fetch
6 ms
03Local LLM
3.4 B params · INT4 · NPU
74 ms
04TTS
streaming · 24 kHz · prosodic
24 ms

142 ms median round-trip

On-device, end-to-end. Faster than the breath you take to speak.

Persistent memory

Survives restart, sleep, silence. The friend who actually remembers.

Turn-aware

Interrupts gracefully, never overlaps. Knows when to wait.

Private by default

Nothing leaves the chip without consent. Cryptographic clone-gate.

NPCs that improvise.

Companions who riff with the player. Narrators who pace a scene. Factions who whisper between themselves. Multimodal, in-engine, royalty-free per session.

<80 msvoice moderation at the edge
Vex · ch.02Did you really come back for me, or for the lantern?
Vex · narrator
Thessa · player

Living NPCs

Memory of every choice, in-character.

Dynamic narration

A director that paces drama in real time.

Voice moderation

Toxic-speech filter at the edge, <80 ms.

Streamer companions

A co-host that reads chat and rolls.

Long talks on moving things.

Hour-long dialogues across noisy floors and shifting context. Agentic voice that asks for clarification, plans the next motion, and reports back to the fleet.

83 dBnoise floor · industrial-grade
UNIT · NRX-04BAY 12 · CORRIDOR EAST
NRX-04 says“On my way. Two minutes.”
83 dB noise floor83dBsurvives industrial floors
Plan length6.4mroute confidence intact
Rotate−38°next motion delta
Plan confidence97%before commitment

Fluent where others forget.

Khaleeji, Levantine, Maghrebi, Egyptian. Amharic, Hausa, Pashto, Sinhala. Forty-one dialects most foundation models cannot place on a map — placed, voiced, understood.

104 / 41languages / low-resource dialects
languages000
low-resource00
Khaleeji ArabicDoha96%
Egyptian ArabicCairo94%
Maghrebi DarijaCasablanca91%
Levantine ArabicBeirut93%
SwahiliNairobi89%
PashtoKabul84%
Hindi · HinglishMumbai95%
MandarinShanghai97%

Why Nuri

The numbers that matter when a voice is between you and a person.

MetricCloud-only LLMOpen-source TTS+ASRNuri
Median end-to-end latency820 ms540 ms142 ms
Languages in production2640104
Low-resource dialects3941
Runs on-devicenopartialyes
Persistent memorystatelessnoyes · local
Interruption / barge-innopartialnative
Voice cloning consentcryptographic

From builders, in production

We shipped a Khaleeji voice agent in nine days. The customers thought it was a person.
Layla Al-MansouriVP Product · Mahara Bank · Doha
0 daystime to deploy
Khaleejizero training data
00%customer satisfaction

By the numbers

Voice that lives everywhere a body moves.

000msmedian end-to-end latency, on-device
000languages and dialects in production
00low-resource dialects from raw field audio
0×smaller than the comparable cloud-only model

Pricing

One pricing page. No surprises in production.

Pay for what you ship, not what we trained. Persistent memory and on-device runtime are not premium add-ons — they are the product.

Studio

Freeforever

Prototype voice in your dialect. Stay private, stay local.

  • 60 minutes of streaming voice / mo
  • On-device runtime · Mac / Linux
  • English + 10 high-resource languages
  • Community Discord
Start free
Most builders

Builder

$0.012/ minute

Ship to production with the full dialect map and persistent memory.

  • 142 ms median round-trip · production
  • 104 languages · 41 low-resource dialects
  • Persistent memory · cryptographic clone-gate
  • On-device + edge fallback
  • 99.9% SLO · email support
Talk to research

Enterprise

Custompriced per fleet

Robotics fleets, regulated voice, sovereign deploys, custom dialects.

  • Dedicated runtime · air-gapped option
  • Custom dialects from raw field audio
  • Hardware partner co-deploy
  • Compliance: HIPAA, SOC2, ISO 27001
  • Named research partner
Plan a fleet

All tiers include cryptographic voice-cloning consent · open weights for academic research · dialect contribution program for low-resource languages

Conversation starter

Begin a conversation
that ends with shipping.

A 30-minute call with our research team. We'll bring a working prototype in your dialect, your domain, your hardware.

✓ Sent. We'll be in touch within a day.