AI tools and news, explained in plain words
Tools

Inworld Voice Isolator: free AI background noise remover

Inworld's Voice Isolator is a free, open-source (Apache 2.0) AI tool that strips noise from voice recordings at 48 kHz, offline or live. How to use it.

Checked and updated 22 November 2026 at 03:55 UTC

Inworld AI logo artwork
Image: Inworld AI

Key points

  • Voice Isolator is Inworld AI's first open-source model, released November 18, 2026: it removes background noise, music and chatter from speech at 48 kHz while keeping breaths, laughs and the speaker's tone.
  • It is free, including for commercial use, under the Apache 2.0 license. You install it with one pip command and it runs on Nvidia GPUs, Apple-silicon Macs or a plain CPU. There is no web app; it is a command-line and Python tool.
  • It comes in two versions: an offline model (64M parameters) for the best quality on recordings, and a realtime model (58M) for live calls with about 63 ms of delay. Inworld's own benchmark ranks both above other open-source denoisers.

Inworld Voice Isolator is a free, open-source AI model that removes background noise, music and other people's chatter from voice recordings while keeping the voice itself, breaths and laughs included. It is released under the Apache 2.0 license, so you can use it for commercial work too, and it runs on your own computer: an Nvidia GPU, an Apple-silicon Mac or even a plain CPU.

Inworld AI, which builds AI models and inference for consumer apps, published it on November 18, 2026 as its first open-source model, and it is trending on Hugging Face this week. Here is what it does, how to use it and where it falls short.

What it does

Most noise removers treat a laugh, a sigh or a breath as noise and cut it. Inworld says Voice Isolator is trained to keep them, so the cleaned voice still sounds like the person. Its demo clips remove a car, a vacuum cleaner, a blender and rock music from speech.

It also saves what it removed as a separate file, which is handy if you want the background track (for example, to separate vocals from music).

Two versions

voice-isolator-v1 voice-isolator-v1-realtime
Best for Recordings: podcasts, interviews, voiceovers Live audio: calls, streams, voice agents
How it works Looks at 8-second chunks at a time Streams with 20 ms of look-ahead
Delay Not live About 63 ms
Size 64 million parameters 58 million parameters
Audio 48 kHz mono 48 kHz mono

Both are small by AI standards, which is why they can run on a laptop.

How to use it

You need Python. There is no website or app yet, but the steps are short (from Inworld's model card):

  1. Install it: pip install inworld-voice-isolation
  2. Clean a file: voice-isolate noisy.wav clean.wav. The weights download from Hugging Face the first time.
  3. You get clean.wav (the voice) and clean_background.wav (everything removed), both 48 kHz WAV files.
  4. Add --no-background to save only the voice, or --output-sr 0 to keep your file's original sample rate.
  5. For live-style processing, use --mode streaming --lookahead 2.

It picks Nvidia CUDA if you have it, Apple's Metal on M-series Macs (PyTorch 2.6 or later), and the CPU otherwise. The GitHub repository also has a local demo app with a file tab and a microphone tab (up to 60 seconds). Inworld has tested it on data-center Nvidia GPUs and CPUs; smaller consumer GPUs are not officially tested yet.

How good is it?

Inworld built its own benchmark, VIBE, which scores from 0 (unprocessed input) to 1 (the clean reference):

Open-source denoiser (48 kHz) VIBE score
Inworld Voice Isolator 0.768
Inworld Voice Isolator realtime 0.718
ClearerVoice MossFormer2_SE_48K 0.695
resemble-enhance (denoise) 0.646
FastEnhancer-L 48k 0.589
DeepFilterNet3 0.510

Keep in mind the benchmark and its test set are Inworld's own and private for now, so these are self-reported results.

Limits to know

  • It won't fix a bad mic. Version 1 removes noise; it doesn't restore missing detail. Inworld says restoration is coming in version 2.
  • It keeps the loudest voice. A quieter person talking over you gets removed; two voices at similar volume usually both stay.
  • Mono only, one channel at 48 kHz.

For more free voice AI, see Haru-4B, an open model that talks and listens at once and our explainer on ZONOS3-Mini.

AI Tools This Week is not affiliated with Inworld AI.

Frequently asked questions

Is Inworld Voice Isolator free?

Yes. The code and weights are released under the Apache 2.0 license, which allows personal and commercial use. You run it on your own computer, so there is no subscription or per-minute fee.

Can Inworld Voice Isolator fix a bad microphone?

No. Version 1 only removes what surrounds the voice (denoising). It does not restore detail a cheap mic or phone line never captured. Inworld says restoration is planned for version 2.

Does it separate two people talking?

Not really. It keeps the foreground voice: if someone a few dB quieter talks over you, they are removed, but two voices at similar volume are usually both kept. It is not a speaker selector.

Sources

  1. Introducing Voice Isolator: open-source 48 kHz denoising (Inworld AI blog, Nov 18, 2026)
  2. inworld/voice-isolator model card (Hugging Face)

The AI Tools This Week Desk writes and edits every article, checks claims against primary sources and updates stories when the facts change.

Related