Crele Focus 1.1: speaker isolation built on an audio world model

What is new in Crele Focus 1.1?

Crele Focus 1.1 gives you a clearer voice and stays on the right speaker from start to finish. It replaces Crele Focus 1 in the Crele app, on every plan.

• One voice, start to finish. It learns what your speaker sounds like in this recording, not only in the sample you gave it, and keeps hold of them through other people’s turns and long pauses.
• Rebuilt, not cut out. Crele Focus 1 removed the other voices. Crele Focus 1.1 rebuilds your speaker’s voice instead.
• Silence over the wrong voice. Where it cannot tell your speaker from someone else, that stretch comes back silent rather than carrying another person’s words.
• No charge for a miss. If it cannot find your speaker, the run uses none of your minutes.
• Shorter waits. Long recordings come back sooner than they did on Crele Focus 1.

Nothing else changes: you give it the recording and three seconds or more of your speaker, on the same plans.

What is an audio world model?

An audio world model, the approach behind Crele Focus 1.1, is a model that learns the sound of the place a recording was made in as a whole: every voice in it, who is speaking when, what each person sounds like, and the room and the noise around them. Crele uses the term for an approach rather than one model. Instead of training a model for a single job, such as removing noise, you train it to understand the scene, and each job draws on that. Crele Focus 1.1 is Crele’s first model built this way.

Three things set it apart from a model built for one task:

• It knows who is there. At every moment it keeps track of which voices are present and which of them are speaking.
• A voice stays the same voice. Someone who goes quiet for ten minutes is the same person when they speak again, and a voice under another voice is still itself.
• It is trained on what is covered. It learns to recover each person’s speech, including where another voice talks over it.

How is an audio world model different from a language model?

A language model learns from text and predicts the next word. An audio world model, like the one behind Crele Focus 1.1, learns from sound as it happens in a place, and keeps track of who is speaking and what each voice sounds like over time.

In AI, “world model” has mostly meant video and robotics: models that learn how a scene behaves, not only what it looks like. The audio version applies the same idea to hearing. A dinner, a meeting or a street interview is a scene with several sources in it, and a model that understands the scene can answer for any one of them.

That matters for voice AI. A voice assistant built on a language model gets speech after it has been turned into words, and by then who said them, and what else was in the room, is gone. An audio world model is the part that hears: the ears of voice AI.

How do you isolate one voice from a long recording with Crele Focus 1.1?

Upload the recording to Crele Focus 1.1 in the Crele app, give it three seconds or more of your speaker talking alone, and it hands back that one voice for the whole length of the file.

1. Upload the recording. Audio or video: from a video, the app takes the audio track. The longest file depends on your plan, and the pricing page has both.
2. Show it your speaker. Mark a stretch of the recording where only they talk, or upload a separate clip of them on their own. Three seconds is enough, and clean matters more than long.
3. Listen, then download. Switch between the original and the result at any point, then save the result as a WAV file.

Because Crele Focus 1.1 learns your speaker from the whole recording, a sample from the first minute is enough to find them in the last.

What does Crele Focus 1.1 do when it cannot tell who is speaking?

Crele Focus 1.1 leaves that stretch silent. Where your speaker cannot be told apart from someone else, the result is silence rather than another person’s words, and a recording your speaker is not in at all is refused without using any of your minutes.

That is a choice. An isolation that guesses wrong is worse than a gap: a sentence from the other side of the table ends up in an edit, a transcript or a clip, and nobody hears it until it is out. A gap shows itself.

If a result has more silence than you expected, look at the sample first. A few seconds where your speaker talks alone, with nobody under them, gives Crele Focus 1.1 the clearest picture of who to keep.

What comes next for Crele’s audio world model?

Crele is taking the same approach to the places a machine has to hear one person among many: real-time speaker extraction, which follows one voice live for voice agents and devices, and a developer API. Both are coming soon.

Questions

Is Crele Focus 1.1 free?

Yes, on a free Crele account, which includes downloads and a monthly allowance of audio. Pro takes longer files and more audio a month. The pricing page has both plans.

Does the free speaker isolation tool on this site use Crele Focus 1.1?

No. The free tool on this site runs Crele Focus 1 on the first 45 seconds of a file, with no account. Crele Focus 1.1 runs in the Crele app, on a free account.

What is the difference between Crele Focus 1 and Crele Focus 1.1?

Crele Focus 1 takes the other voices out of the recording and hands back what is left, room and all. Crele Focus 1.1 rebuilds the chosen voice from the recording, follows it through the whole file, and leaves a stretch silent where it cannot tell that voice from another.

What does "voice not found" mean in Crele Focus 1.1?

It means Crele Focus 1.1 could not find the voice in your sample anywhere in the recording, so it stopped rather than hand back a silent file. The run uses none of your minutes. Try again with a sample where that person speaks alone.