A painted night sky where the stars trace a speech waveform above a small observatory
CelestLabs · frontier voice AI

Deploy frontier voice AI anywhere.

feather and starling run on ordinary CPUs, so speech and voice cloning go where GPUs can't: every cloud region, your own datacenter, air-gapped, wherever your data has to stay. From $1.90 per million characters, a fraction of GPU-bound APIs.

PRESS PLAY · STARLING · 9 VOICES · 4 LANGUAGES
"Hear the same engine take a support call in English, read a bedtime story in Hindi, confirm a booking in Arabic, and deliver in Spanish."
CELESTLABS · SAN FRANCISCO
37.77°N 122.42°W · EST. 2026
SERVING · FEATHER + STARLING
FEATHER
382ms
to the first word · CPU only, measured
STARLING
107
languages · studio quality · cloning
PRICE
$1.90/1M
characters on feather · $9/1M starling
CLONING
30s
of audio to a production clone
01 · Two Models, One API

Pick your instrument.

Same OpenAI-compatible endpoint, two very different price-performance points. Mix them per request: feather for the high-volume paths, starling where the voice IS the product.

FOR SCALE · SERVING
feather
The workhorse. Real-time speech on ordinary CPUs, priced to leave on for every user.
First word382 ms · measured
Throughput6.1× realtime · CPU
Languages23
Streamingsentence-by-sentence
Deploysevery region · on-prem
$1.90 / 1M characters
FOR QUALITY · SERVING
starling
The flagship. Studio-grade voices, expressive delivery, and cloning, in 107 languages.
Languages107 · all Indic
Voice cloning30 s reference
Expressionlaughs · whispers · sighs
Qualityevery clip on this page
DeploysCPU · no GPUs
$9 / 1M characters
107 languages, one voice

A single clone, or any preset voice, speaks every language below, including scripts the original speaker never learned. Every major Indian language is covered.

EnglishHindiSpanishMandarin ChineseArabicBengaliPortugueseRussianJapaneseFrenchGermanUrduIndonesianItalianTurkishTamilTeluguMarathiVietnameseKoreanThaiGujaratiPolishUkrainianMalayalamKannadaPunjabiPersianRomanianDutchGreekCzechSwedishHungarianHebrewTagalogMalayBurmeseNepaliSinhalaKhmerLaoOdiaAssameseAmharicSwahiliYorubaIgboHausaZuluAfrikaansSomaliDanishFinnishNorwegianSlovakBulgarianCroatianSerbianSlovenianLithuanianLatvianEstonianGeorgianArmenianAzerbaijaniKazakhUzbekMongolianPashtoKurdishSindhiTibetanUyghurCantoneseJavaneseSundaneseCebuanoIcelandicIrishWelshBasqueCatalanGalicianMalteseLuxembourgishAlbanianMacedonianBosnianBelarusianTatarTurkmenKyrgyzTajikDariWolofXhosaShonaKinyarwandaMalagasyHaitian CreoleMaoriSamoanFijianHawaiianBhojpuriKonkani
02 · Voice Cloning

Any voice you own, yours in 30 seconds.

starling builds a production-ready clone from half a minute of clean audio: your brand ambassador, your founder, your support lead. The clone speaks all 107 languages, even ones the speaker never learned.

Record 30 seconds. A phone memo is enough; a quiet room is better.

Upload once. The clone is frozen as a private voice on your account.

Speak everywhere. Same voice, any script, any of 107 languages, via one API call.

clone · api.leanvoice.aiSTARLING
# one call: reference in, cloned speech out
curl https://api.leanvoice.ai/v1/audio/speech \
  -d '{ "model": "starling",
       "ref_audio": "@you_30s.wav",
       "language": "hi",
       "input": "नमस्ते, मैं आपकी क्या मदद कर सकता हूँ?" }'
 
✓ cloned voice · speaking Hindi
Clone Your VoiceCloning Docs
03 · Voices In The Wild

Drop a voice into any room.

Real products aren't used in a silent booth. starling can place a voice inside a room, a cafe, a call floor, a moving street, so agents, IVR, and announcements sound like they belong there. One request parameter, no post-production. Press play on any room.

Your own roomambient_noise + levelPass any bundled bed, or upload your own, and set how loud it sits under the voice.
speech · api.leanvoice.aiSTARLING
# same call, now it sounds like a real cafe
curl https://api.leanvoice.ai/v1/audio/speech \
  -d '{ "model": "starling", "voice": "meera",
       "input": "Your oat milk latte is ready.",
       "ambient_noise": "cafe", "ambient_level": 0.3 }'
An aurora rendered as a sound spectrum over dark mountains
04 · The Stream

Fast enough for a conversation.

One OpenAI-compatible call. Sentences stream while the rest is still being synthesised, so voice agents answer like people, not like voicemail. 382 ms to the first word; 6.1× faster than playback for everything after it.

Read the streaming docs

05 · The Bill

Seven engines. One sentence. Watch the bill.

Each row is a real recording of the same line from that provider's model. The meters count up in lockstep with playback: each number is that model's price per million characters, not the cost of this clip.

FIG. 01 · seven engines, published list prices per 1M characters, each paired with the model you hear.

CelestLabsstarling · studio
$0.00/ 1M chars1× · you
Smallestlightning-v3.1
$0.00/ 1M chars1.9×
Deepgramaura-2
$0.00/ 1M chars3.3×
OpenAItts-1-hd
$0.00/ 1M chars3.3×
Sarvambulbul:v3
$0.00/ 1M chars3.5×
Cartesiasonic-3.5
$0.00/ 1M chars4.4×
ElevenLabseleven v3
$0.00/ 1M chars11×

Published list prices per 1M characters · verified June 2026

Earth from orbit at night: a dense web of compute lights across every continent, with only a few blazing points of concentrated GPU power
06 · Everywhere

Deploy where your users actually are.

feather and starling run on ordinary CPUs, so they go everywhere CPUs already are: every cloud region on earth, your own datacenter, even air-gapped. The GPUs behind frontier AI don't. That gap is your data-residency story, and a moat GPU-bound vendors can't cross.

The map of scarcity

CPUs run in every region. The best GPUs run in a third.

Every major cloud offers general-purpose CPUs in 100% of its regions. The H100 and A100 GPUs behind frontier models reach only a fraction, so a GPU-bound vendor cannot serve customers whose data has to stay in the other two-thirds.

AWS
39 regions
CPU compute39 / 39
H100 / A10015 / 39
Azure
60+ regions
CPU compute60 / 60
H100 / A100~17 / 60
Google Cloud
43 regions
CPU compute43 / 43
H100 / A10019 / 43

Total regions versus H100 and A100 availability, mid-2026. Even counting every inference GPU (T4, L4, A10G), coverage stays partial: roughly 64% of AWS regions, 40% of Azure, 77% of GCP. CPUs are in 100% of all three. Sources: AWS, Azure, Google Cloud.

Moonlit terraces shaped like waveforms
07 · The Economics

Priced to be left on.

No rented GPUs in the path, so the rates hold in every cloud region, on-premise, and air-gapped.

feather · per 1M characters$1.90
starling · per 1M characters$9.00
feather per minute of audio$0.0017 · 15 paise
Free, renewed monthly10,000 chars
Full pricing
CelestLabs

Give your product a voice tonight.