CelestLabs · Research

Frontier voice, on ordinary CPUs.

What we measured and what we ship. Every number on this page is taken from the model itself: parameter counts read from the shipped ONNX graphs, timings from a warm CPU at batch one.

MODELS · FEATHER · STARLING · SERVING  ·  METHOD · FLOW MATCHING, 8-STEP  ·  COMPUTE · CPU ONLY
OBS. 01
<100ms
feather · time to first word
OBS. 02
6.1×
faster than realtime · CPU only
OBS. 03
0.18
real-time factor · warm, batch one
OBS. 04
107
languages across the lineup
01 · The model

Model card: feather.

Four neural networks in sequence. Text is encoded; a duration model decides how long each sound lasts; a flow-matching field sculpts random noise into a mel-spectrogram in eight fixed Euler steps; and a vocoder turns that spectrogram into a 44.1 kHz waveform. The field is the only iterative stage; everything else runs once.

SpecReadingValue
generative coreConditional flow matching along a near-straight transport path64.0M
solverFixed eight-step Euler · eight network passes per utteranceNFE 8
vocoderMel-spectrogram to 44.1 kHz waveform25.3M
encoder + durationRun once per utterance · text and timing9.9M
parametersFull model 117M · shipped inference graphs 99.2M, measured117M
speedWarm, batch one, on a commodity core · no GPU in the pathRTF 0.18
first wordStreaming fast path · opener cached, gates overlapped<100 MS
languages107 across the lineup · one voice registry, one API107
voicesSix presets on feather · starling adds zero-shot cloning from under 30 s of audio6 + CLONING
computeOrdinary CPUs, x86 and Arm · no GPU anywhere in the path● CPU

Built on open foundations: an open flow-matching speech model with MIT-licensed inference code and Open RAIL-M weights, which we export to ONNX, optimize, and serve. Our contribution is the inference and serving stack; attribution is retained per those licenses. The remaining ~18M parameters are a style encoder used offline to build voices, never run at synthesis time.

field measurement · laptop cpuBATCH = 1
# single request · warm · eight steps
audio produced   4.03 s · 44.1 kHz mono
compute time     0.72 s
rtf             0.18 · well under realtime
first word       <100 ms · streaming
ode steps        8 · fixed Euler
gpu required     none
 
✓ headroom one core carries several realtime voices

Most diffusion systems need tens to hundreds of network passes per utterance: roughly ~1,000 for high-quality diffusion, ~50 for fast variants. Because the flow path is nearly straight, eight suffice. That, plus a network small enough to keep its weights in cache, is the whole trick.

A real spectrogram plotted as a star chart
FIG. 01 · "every voice is a constellation," spoken by meera, plotted star for starREAL MODEL OUTPUT

Not an illustration: a real utterance, its spectrogram plotted star for star. Hear every voice in the voice gallery, or synthesise your own in the dashboard.

02 · The measure

Speed becomes price.

An RTF of 0.18 on a commodity core is the whole business case. A GPU serving realtime voice runs at batch one, its worst case, and idles most of its silicon; a CPU at batch one is doing exactly the work it is good at. The floor that collapses is the price: feather at $1.90 per million characters, 0.19¢ per 1,000, and it holds in every region, inside your own cloud, on-premise, and air-gapped.

Provider$/1K chars$/min audioCompute
CelestLabs feather$0.0019$0.0017CPU
CelestLabs starling$0.0090$0.0081CPU
OpenAI tts-1$0.015$0.015GPU
Amazon Polly$0.016$0.016GPU
Cartesia Sonic$0.036$0.036GPU
ElevenLabs$0.150$0.100GPU

At list prices, feather is roughly 8× cheaper than Polly, 19× cheaper than Cartesia, and up to 50× cheaper than frontier voice models. Per minute of audio: $0.0017, about 15 paise; per hour, $0.10. Competitor rates are published list prices and move; starling per-minute is derived from the same characters-per-minute assumption as feather. Full detail on the pricing page.

03 · The lineup

Two models, one API.

Both instruments are serving. feather is the cheap, fast workhorse for agents and high-volume narration; starling is the expressive flagship for cloning, media, and lifelike agents. Same endpoint, same keys, one line of difference in the request.

ModelNotesStatus
feather-tts-1Flow matching · 8-step · 117M · RTF 0.18 · under 100 ms to first word · $1.90 per 1M characters● SERVING
starling-tts-1Zero-shot voice cloning · 107 languages · narration and media · $9.00 per 1M characters● SERVING

Both run on ordinary CPUs, so either model deploys in any region, inside your own cloud, or fully on-prem. Neither exposes an emotion or expression control: register comes from the voice you pick, the punctuation, and the phrasing.

Keep reading

Put it to work.