ElevenLabs Audio Isolation: clean speech out of a noisy recording

An interview recorded next to a road. A voice memo with a fan running. A video where the wind ate half the words. ElevenLabs Audio Isolation is for exactly these: you upload the recording, it keeps the voice and removes what is around it. There is no prompt and there are no settings. That makes it the easiest model on the models page to use and the easiest to misuse, so this post is mostly about what to feed it.

What it is good at

• Steady noise: air conditioning, fans, hum, traffic, the hiss of a cheap microphone.
• Busy rooms: cafe chatter, an office behind you, a crowd at an event, as long as your voice is the loudest one.
• Wind and handling noise on phone and camera recordings.
• Rescuing dialogue before you dub, subtitle or edit it, since every later step works better on a clean voice.

How to run it

1. 01Open the model and upload one audio file. The studio takes the common formats: MP3, WAV, OGG, M4A and FLAC.
2. 02Check the price on the Generate button and run.
3. 03Download the cleaned file and listen to it on headphones, not laptop speakers, before you use it.

Lifehacks

• Cut before you clean. Trim the silence at the start and the end and the parts you will not use. A shorter file is quicker and cheaper, and you only pay to clean what ships.
• Upload the best copy you have. A WAV or a high-quality original holds more of the voice than an MP3 that has already been squeezed by a messenger app. The model can only keep what is there.
• Mix a little room back in. Perfectly dry speech can sound pasted on. In your editor, lay a quiet bed of the original under the cleaned track, or a soft music bed, and the voice sits naturally again.
• Test with the worst minute. Before cleaning a long recording, cut out its noisiest stretch and run that alone. If the worst part comes out usable, the rest will.

Who should use something else

This is a voice keeper, not a mixing desk. If you want to keep the music and remove only the singer, or split a song into parts, look for a separation model on the models page. If you need new sound for a silent clip, a video to sound or a text to sound effects model is the right tool. And for generated clips, many video models now write their own sound in the same run; the video model guide covers which do, and the photo to video guide shows where that fits.

The honest limitation

It cannot bring back what the noise destroyed. When a truck drives past at the exact moment a word is spoken, the word comes out thin or half missing, because it was never clearly recorded. The same goes for distortion: a voice that clipped because the microphone was too close stays clipped. And it treats everything that is not speech as noise, so a guitar played in the background of a podcast leaves along with the hum. Record as close and as clean as you can, and use isolation for what you could not avoid.