Audio AI that works in the real world.
Real recordings have a second person talking, a TV on in the background, a café table too close. We build models for that audio, not for clean input.
A real-time audio decision model. Ask yes/no questions about live audio in plain English and get an answer every 40 ms, including whether the user or someone else is speaking. Runs on your own machine. Apache-2.0.
pip install git+https://github.com/Crele-AI/nimbus
python -m crele_nimbus.playground --openWeights on Hugging Face · Model page · Release post
A benchmark for keeping one voice from a conversation, given a short sample of that voice. Every score comes from the output audio alone, so any file-in, file-out system can be tested, closed products included.
Dataset on Hugging Face · Results · Release post
- Speaker Isolation, Crele Focus 1.1. Mark one voice in a recording and get it back alone, every other voice removed.
- Speech Enhancement, Crele Remaster 1. Upload one voice or a whole conversation and get it back without the noise or the room. Remaster regenerates the voice rather than filtering it.
- Speech Editing, Preview. Change the words in a recording and hear them back in the speaker's own voice.
Create a free account to use all three.
- Real-time Speaker Extraction. Enroll 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. Our models as endpoints from your own code, then the live stream. Tell us what you are building.
Website · Blog · Hugging Face · LinkedIn · Email us