The Best Free and Local AI Stem Separation Tools in 2026

Key Takeaways

• Top pick: Ultimate Vocal Remover with Roformer models. It's the best free quality you can run on your own machine, with a real installer for Windows and Mac.

• Best maintained: python-audio-separator. Runs the same Roformer models from the command line, and it shipped three releases in August 2026.

• Demucs is still the four- and six-stem workhorse. It's MIT-licensed, sits inside most of the tools on this list, and runs on a CPU at roughly 1.5 times the length of the track.

• The very best scores aren't downloadable. MVSep's top vocal model, its own BS-RoFormer build, scores 12.33 dB SDR on its benchmark, and we found no download for it. The best public weights you can run locally sit around 11.35.

• The license covers the software only. Splitting someone else's track doesn't give you the right to release anything made from it.

Why run stem separation on your own machine

If you own a recent DAW, start there. Logic Pro splits six stems on the device, and Ableton, FL Studio, Cubase and Fender Studio Pro all have their own separators now, in some editions; we covered them in our roundup of AI audio editing tools. This piece is for everyone else: producers on a DAW without a splitter, remixers who want better vocals than their DAW gives them, and anyone who doesn't want to upload unreleased mixes to a web service.

The free, local options are better than you'd expect, because the strongest separation models of the last two years are open. ByteDance researchers published the BS-RoFormer and Mel-Band RoFormer architectures in 2023 without releasing weights, and community trainers built their own. On MVSep's comparison table, the best public Mel-Band RoFormer weights score 11.35 dB SDR (signal-to-distortion ratio, higher is better) on vocals, against 8.18 for Meta's default Demucs model. Every tool below is a different way to run those models, or Demucs, without paying anyone.

How we picked

Four criteria.

• It runs offline. Nothing gets uploaded, and there's no account or queue. That ruled out MVSep's own site, which hosts the top-scoring models but processes them on its servers.

• It's free, and the license says so. That ruled out Open-Unmix's best model, whose weights are licensed for non-commercial use only. Spleeter is free and MIT-licensed but scores 5.9 dB overall SDR against 9.0 for fine-tuned Demucs, so it didn't make the cut.

• Someone still ships it. We note the last release for each tool, because several are effectively in maintenance mode.

• Quality is measurable. We lean on MVSep's public Multisong benchmark. SDR measures how close a stem is to the original waveform. It doesn't measure how the stem sounds, so use it to compare models.

Nobody at Studio Desk ran a listening test across these tools. The ranking comes from each project's own documentation and release history, MVSep's benchmark, and the GitHub issue trackers linked below, all checked on October 1, 2026.

At a glance

1. Ultimate Vocal Remover - the best free quality, with a real installer

Ultimate Vocal Remover (UVR) is the free desktop app most people mean when they talk about local stem separation. It runs a library of models (MDX-Net, MDX23C, Demucs and the VR architecture) that you download from inside the app, and the Roformer beta builds, released between November 2024 and January 2025, add Roformer models (BS-RoFormer and Mel-Band RoFormer among them) plus SCNet and Bandit, and the Roformers are where the quality is. There are installers for Windows 10 and later and separate Apple silicon and Intel builds for macOS Big Sur and later. Linux users install from source.

Hardware matters. The project recommends an Nvidia card with 8GB of VRAM (a GTX 1060 with 6GB is the floor), and AMD and Intel Arc cards get a DirectML build with limited support. On Apple silicon, GPU acceleration covers Demucs v4 and MDX-Net. Roformer models on a CPU are slow enough to leave running while you make dinner.

The bigger concern is upkeep. The last stable release, 5.6, is from September 2023, the last code commit is from October 2023, and the issue tracker holds about 1,500 open reports with little visible triage. One open issue says Roformer got several times slower in the January 2025 beta. It still works. Just don't expect fixes.

Best for: Producers who want the best free vocals and instrumentals without touching a terminal.

Price: Free, as of October 2026.

Platforms: Windows, macOS; Linux from source.

License: MIT for the app. Third-party model weights carry their own terms.

2. python-audio-separator - the same models, actually maintained

python-audio-separator runs the models from UVR's ecosystem, including BS-RoFormer and Mel-Band RoFormer, Demucs and MDX, from the command line or a Python script. Version 0.47.0 shipped on August 27, 2026, after releases that added half-precision inference and faster Roformer processing on Apple silicon. It installs with pip in CPU, CUDA or DirectML flavors, uses Core ML on Apple silicon, and downloads models the first time you call them. The default is a BS-RoFormer vocal model.

There's no GUI, and you need Python 3.10 or later. If a command like audio-separator song.wav means nothing to you, use UVR instead. If it does, this is the better-maintained way to run the same weights, and it scripts well for batch jobs.

Best for: Producers comfortable in a terminal, and anyone splitting a whole folder of tracks.

Price: Free, as of October 2026.

Platforms: Windows, macOS, Linux.

License: MIT. Individual model weights vary.

3. Demucs - the four- and six-stem workhorse

Demucs is the model Meta's research team released in 2019 and kept improving through version 4, the Hybrid Transformer release. Its default htdemucs model splits vocals, drums, bass and other; htdemucs_6s adds guitar and piano. The original repository was archived in January 2025, and important bug fixes continue at the author's own fork, which shipped version 4.1.0 in July 2026 with lighter dependencies and models hosted on Hugging Face.

It installs with pip and runs with one command. The project says it needs 3GB of GPU memory, 7GB recommended, and processes on a CPU in about 1.5 times the track's length. The README is honest about the six-stem model: guitar is okay, piano has a lot of bleeding and artifacts. For vocals specifically, the Roformer models in the two entries above score about three dB higher on MVSep's benchmark.

Best for: Drums, bass and full four-stem splits, and anyone building their own tools.

Price: Free, as of October 2026.

Platforms: Windows, macOS, Linux.

License: MIT for the code. The pretrained models ship from the same project with no separate license stated.

4. StemDeck - the friendliest six-stem app, still in alpha

StemDeck is a new open-source desktop app that wraps Demucs' six-stem model in a clean interface, with a built-in mixer for soloing and muting stems plus BPM and key detection. It has downloads for Windows, macOS and Linux, picks up an Nvidia or Apple silicon GPU automatically, and falls back to the CPU.

Bedroom Producers Blog called it still-alpha software in August 2026, and at version 0.19 it's still pre-1.0, so that's the right expectation. It also runs only Demucs, so its vocals are a step below what UVR and audio-separator get from Roformer models. For a quick practice split or a first look at an arrangement, it's the easiest thing here to install.

Best for: Musicians who want six stems and a mixer without any setup.

Price: Free, as of October 2026.

Platforms: Windows, macOS, Linux.

License: Apache-2.0.

5. Audacity OpenVINO plugin - separation inside a free editor

Intel's OpenVINO plugins for Audacity add a Music Separation effect that runs Demucs v4 locally on your CPU, GPU or NPU. It splits two stems (vocals and instrumental) or four, and a Shifts setting trades processing time for quality. If you already edit in Audacity, it puts separation one menu away.

The limits are platform and age. There's a Windows installer, Linux users build from source or use the Audacity snap, and there are no macOS instructions. The latest release targets Audacity 3.7.1 and dates from December 2024. Quality is Demucs quality, so it's fine for drums and bass and behind Roformer on vocals.

Best for: Audacity users on Windows who want separation without another app.

Price: Free, as of October 2026.

Platforms: Windows; Linux from source or snap.

License: GPL v3.

6. Music Source Separation Training - for chasing the newest weights

Music Source Separation Training (MSST) is the toolkit the MVSep team maintains for training and running separation models, and it's where many of the community Roformer weights first appear. Its pretrained model list links dozens of checkpoints for vocals, drums, bass and other instruments, and an inference script plus a basic GUI run them locally.

It's a research toolkit. You're expected to manage Python environments and config files, and the documentation assumes you know what a checkpoint is. The pretrained list doesn't state licenses for the weights, so check each one before using output commercially.

Best for: Technical users who want a new fine-tune the week it's released.

Price: Free, as of October 2026.

Platforms: Windows, macOS, Linux.

License: MIT for the code. Weight licenses are unstated.

What the license does and doesn't cover

Every license above covers the software and, sometimes, the model weights. None of them gives you rights to the recording you feed in. The U.S. Copyright Office treats a sound recording and the composition in it as two separate works, and both owners hold the exclusive right to prepare derivative works. In practice, a remix built from someone else's stems generally needs permission from both. Separating stems for practice, transcription or study is a different question from releasing a remix or a sample. This isn't legal advice; if you plan to release something, clear it first.

Which one fits your desk

• You want the cleanest vocal or instrumental, no terminal: UVR's Roformer beta, on a machine with a decent GPU.

• You're splitting a back catalog or scripting a workflow: python-audio-separator.

• You need drums and bass, or six stems for practice: Demucs from the command line, or StemDeck if you want a mixer and no setup.

• You edit in Audacity on Windows: the OpenVINO plugin.

• You already own Logic Pro or Ableton Live Suite: try the built-in splitter first, and reach for this list when it isn't enough.