AI that hears
the real world
Crele builds audio AI for real rooms, real conversations, and real life. Experience our models for free today, more coming soon.
Control audio,
one model at a time
Isolate a speaker or clean up a recording now, free in your browser. Speech editing is in the app as a preview. A live engine and an API are on the way.
Speaker Isolation
Crele Focus 1.1Mark one voice in a recording and get it back alone, every other voice removed. Crele Focus 1.1 is a target speaker extraction model.
Speech Enhancement
Crele Remaster 1Upload a recording, one voice or a conversation, and get it back without the noise or the room. Crele Remaster 1 regenerates the voice rather than filtering it.
Speech Editing
PreviewChange the words in a recording and hear them back in the speaker's own voice. Try a preview of the model today.
Real-time Speaker Extraction
Coming soonEnroll one voice and get it back alone from a live stream, with a signal for when they are talking. For voice agents and devices in busy rooms.
Developer API
Coming soonBoth models as endpoints from your own code, then the live stream. Tell us what you are building.
You can pick one voice out of a crowded room.
AI still can't.
Every voice pipeline runs on a speaker-agnostic front end: it hears that someone is speaking, never who. So agents answer the TV, take orders from bystanders, and stall on background chatter. Diarization numbers the voices. It never says which one is yours, and it breaks down on cross-talk and overlapping speech. In production, other people cause more failures than noise.
Live transcript
Right, so with the extra charger and the case, what's the total coming to?
…and that's tonight at nine, only on Channel 5, right after the late news.
Sorry, could you read that back to me? I think you missed the last item.
I'll take two of whatever he's having, and a large coffee as well.
Every word is right. Two of the speakers are wrong. Nothing downstream can tell.
Mouth
Voice generation. Speech so lifelike you can't tell.
Brain
Understanding and reasoning over what was said.
Ears
Hearing a real, messy scene and knowing who is who, at every moment.
Enroll a voice in seconds. As audio streams, the engine locks onto that identity and returns two things, live: the isolated audio of your speaker with every other voice and sound suppressed, and attribution signals telling you whether they're talking, with what confidence, and whose turn it is.
Maya's Voice
maya-sample.wav
EnrolledJack's Voice
jack-sample.wav
EnrolledEngineered for real audio
Phone lines, equally loud interfering voices, overlapping speech: the conditions that break everything else are the design target.
const session = crele.stream({ voice: enrolled })
session.on("audio", chunk => {
sink.play(chunk) // your speaker, isolated
})
session.on("speaker", state => {
state.speaking // true
state.confidence // "high"
state.turn // "yours"
})A simple streaming API
Isolated audio and attribution arrive over one stream, with drop-in integrations for the tools builders already use.
Voice isolation and attribution come from one engine reasoning about one identity, so they can never disagree. It replaces the front end of every voice pipeline: the part that decides what the rest of the system gets to hear.
Demos are quiet. The world is not.
Voice identity is infrastructure, not a point tool. The same front end sits in front of all six of these, and each is a room where people talk over each other.
Voice agents
Phone, drive-thru, kiosk and in-app agents that answer one customer and ignore the queue behind them, the radio, the passenger seat.
Contact centers
Hold music, cross-talk, a floor of other agents. QA needs to know who said what.
Meetings & dictation
In an open office your voice is the command and every other voice is content.
AI hardware & wearables
Glasses, pendants and always-on devices that wake for their owner and stay shut through the television and the person on the next bench.
Media & recording
Podcasts, interviews and field recordings where every speaker arrives on a track of their own, already cleaned and attributed.
Privacy & compliance
Capture the people who consented. Keep everyone else out of the recording.
Full control of speech and audio
Speech is humanity's native interface, yet audio is the least controllable medium we have. We can edit text character by character and images pixel by pixel, but sound arrives as one undifferentiated wave: every voice, every noise, every room fused together. We exist to change that.
Perceive
Decompose any real acoustic scene: which sounds and voices exist, live.
Attribute
Bind every voice to an identity, frame by frame.
Transform
Translate, restyle and enhance speech while the speaker stays themselves.
Verify
Know that a voice is real, present, and who it claims to be.
Every one of these begins with the same question: whose voice is this? That is why identity comes first, and why everything built on top of it gets better once it exists.
Voice AI has crossed into production at scale, and production audio is messy in ways demos never are.
The stack is shifting toward real-time, speech-native models. That makes the front of the pipeline matter more, not less.
The scientific ingredients for speaker-aware audio have matured. What's missing is neutral, production-grade infrastructure.
Upload a recording now
Sign up for free and run both models on your own recordings, with longer files and downloads: Crele Focus 1.1 to isolate a speaker, Crele Remaster 1 to clean up a recording.
