Drop in any audio or video. Get back the voices, the music and the effects as separate tracks. Mute, solo, add effects, and export. Completely free.
Free forever. No account, no subscription, no catch.
A walkthrough of installing, separating, mixing and exporting.
Built for filmmakers, editors and sound people who got one finished audio track and need it in pieces.
Automatic separation into stems. Everything left over is kept too, so the tracks always add back up to your original.
Mute, solo and set the volume of each track while the video plays. Preview before you export anything.
Pull the Voices track apart into individual speakers, so you can work on one person at a time.
EQ, filters, ring mod, chorus, distortion, delay, reverb, and presets like Robot Voice, Old Radio, Telephone and Cave. Non-destructive.
Export the current mix, a video with the new mix put back in, or every stem as its own file.
Run locally on your own computer, or use a cloud engine. The app checks your hardware and only shows what can run.
Control it from AI apps that support MCP: Claude Desktop, the ChatGPT desktop app and Codex. Just ask: "remove the music from my video." Originals are never changed.
Setup instructions
| Engine | Runs | Good for | Needs |
|---|---|---|---|
| BandIt v2 | On your computer (CPU) | Voices / music / effects. The default. | Nothing extra |
| BandIt v2 (Cloud) | MVSEP servers | Same model, runs anywhere | Free MVSEP account and token |
| DnR v3 (Cloud) | MVSEP servers | Cinematic material. Best results on our hard test clip. | Free MVSEP account and token |
| SAM Audio | On your computer (GPU) | Meta's open model | NVIDIA GPU with about 10+ GB VRAM, plus approved access to Meta's weights on Hugging Face |
Cloud engines upload your audio to a third-party service (MVSEP). Don't use them for anything confidential. Local engines never leave your computer.
No payment, no sign-up. Installers are hosted on GitHub. Pick your system.
Windows 10 or 11 (64-bit). Installs for your user account, no admin rights needed.
Download for WindowsApple silicon or Intel. Downloads as a .zip: unzip it, then open the app. SAM Audio needs an NVIDIA GPU, so on a Mac use BandIt v2 or the cloud engines.
Download for MacRead it, build it, change it, or run it straight from Python. Open source under the MIT license.
View on GitHubUn-Mix-a-Lot is free and new, and I haven't paid for the code-signing certificates that remove these warnings. The code is open source, so you can read all of it on GitHub.
Windows: when you run the installer, Windows may show a blue "Windows protected your PC" screen. Click More info, then Run anyway.
Mac: if macOS says it can't verify the app, right-click (or Control-click) the installer, choose Open, then click Open again. If it still blocks you, go to System Settings → Privacy & Security and click Open Anyway.
Your privacy: your audio stays on your computer unless you choose the MVSEP cloud option. The app also checks GitHub once a day for updates, and you can turn that off in Settings.
Use Un-Mix-a-Lot inside ComfyUI. Three nodes split audio into voices, music and effects, split voices into one track per speaker, and mix the pieces back together. This is an early version, tested on a few setups. Bug reports are welcome.
Splits any audio into three tracks.
One track per speaker. Works best with different people or AI voices. A text box lets you fix mistakes.
Puts the pieces back together. Set a part's volume to 0 to remove it.
In ComfyUI-Manager, choose Install via Git URL and paste the repo address, then restart ComfyUI. Or clone it into ComfyUI/custom_nodes.
The first run downloads a 447 MB model, and after that it works offline. It uses the PyTorch that ComfyUI already has.
Beta software, provided as-is and used at your own risk. See the disclaimer below.
Not affiliated with or endorsed by Comfy Org. ComfyUI and the ComfyUI logo are trademarks of Comfy Org.
Credits: voice splitting uses SpeechBrain (Apache-2.0) and Silero VAD (MIT); the splitting model is BandIt v2 (code Apache-2.0, weights CC BY-SA 4.0).
Yes. No ads, no account, no subscription. The only possible cost is if you choose a cloud engine, where MVSEP has its own free tier and its own limits.
Not with the local engines (BandIt v2 and SAM Audio). The cloud engines upload your file to MVSEP to be processed. Keep that in mind for private or unreleased material.
SAM Audio is Meta's model. Its weights are gated: you need a free Hugging Face account and approval from Meta for each model size. Un-Mix-a-Lot does not include the model, you download it yourself under Meta's license. You also need an NVIDIA GPU with plenty of VRAM. If you don't have one, the app hides it and you can use the other engines.
The separation models work on mono 48 kHz audio, so stereo files are downmixed first and exports are mono.
Un-Mix-a-Lot's own code is MIT licensed, but the models and services it can call have their own terms (see Licenses and credits below). You are responsible for having the rights to the audio you process and for following each model's license.
Open an issue on the GitHub page. This is a free project, so there is no guaranteed support or response time.
Un-Mix-a-Lot is free software provided "as is", with no warranty of any kind. By downloading, installing or using it you agree to the following.
To the maximum extent permitted by law, the software and this website are provided "as is" and "as available", without warranties of any kind, express or implied, including merchantability, fitness for a particular purpose, and non-infringement. In no event shall the authors, copyright holders or contributors be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the software or its use.
Un-Mix-a-Lot's own code is released under the MIT license. It is built on and can use the work of others, who deserve credit and whose terms apply when you use their parts:
See the project's README on GitHub for the full and current list.