When Deezer released Spleeter in 2019, it was a legitimate moment for music production. Tools that could separate vocals had existed before, but most were slow, expensive, or produced output too degraded to use. Spleeter was fast, free, open-source, and ran on consumer hardware. It changed what producers assumed was possible with off-the-shelf software.

Six years later, the landscape has shifted. Spleeter is still around. It still works. But whether it makes sense for you depends a lot on what you’re actually trying to do.

What Spleeter is

Spleeter is a Python library developed at Deezer Research, and you can find the source code and full documentation on GitHub. It supports three separation modes: 2-stem (vocals and accompaniment), 4-stem (vocals, drums, bass, other), and 5-stem, which adds piano as a separate output. The underlying architecture is a U-Net convolutional neural network trained on a large proprietary dataset.

Installation is straightforward for anyone comfortable with Python. pip install spleeter gets you most of the way there, and the command-line interface is clean and consistent. For a developer setting up a batch processing pipeline, it’s one of the easier tools to integrate into an automated workflow.

Where it still holds up

Speed is the main reason to still reach for Spleeter in 2025. It’s fast, especially on CPU, and the memory requirements are modest compared to newer models. If you’re running a pipeline that needs to process hundreds of tracks and you don’t have dedicated GPU capacity, Spleeter’s performance profile is genuinely competitive for the hardware demands.

The 4-stem output is fine for rough work. Pulling a chord chart from an isolated instrument stem, making a practice track, or getting a quick sense of what’s in an arrangement, Spleeter handles all of that without drama. For years it was the default recommendation in this space, and that wasn’t arbitrary.

It’s also well-documented and stable. The codebase hasn’t had major changes recently, which actually means it behaves predictably. There’s no risk of an update changing your output behavior mid-project, and the large body of tutorials and forum posts covering Spleeter makes troubleshooting relatively painless.

Where it shows its age

On quality, Spleeter has been left behind. This isn’t a criticism so much as a simple observation about a field that has moved quickly. HTDemucs, the current version of Meta’s Demucs model, benchmarks significantly better on standard evaluation datasets. The gap on MUSDB18 metrics is substantial.

In practice that gap shows up most obviously as bleed. The bass stem on a dense mix will often contain audible kick drum bleed and melodic leakage. The vocal stem on tracks with busy arrangements will have more instrument spillover than you’d get from a modern model. Acoustic material and jazz, where instruments share a lot of frequency content, make the limitations more pronounced.

Spleeter also doesn’t support 6-stem separation, which breaks out guitar and piano in addition to the standard 4 stems, and there’s no active development pushing quality improvements. What you install today is essentially what the tool has been for several years. For a detailed look at how modern models compare, this post on Demucs, MDX-Net, and HTDemucs covers the evolution well.

Installation and setup reality

For non-technical users, Spleeter isn’t the right tool regardless of quality considerations. The command-line interface, Python dependency management, and occasional environment conflicts add up to more friction than most people want to deal with when they just need stems.

For developers, setup is reasonable. Python 3.7-3.9 works best, as newer Python versions have had some compatibility issues at various points. The pip install usually goes smoothly, and the main gotcha is ffmpeg as a system dependency, which needs to be on the system path. Most developers will sort this out quickly, but it’s worth knowing upfront.

Output quality in practice

On pop and R&B, Spleeter’s 4-stem output is still usable, but it’s noticeably not at the level of what you’d get from a modern tool running HTDemucs or an MDX-Net-based model. The vocals are serviceable. The drums and bass separation is rougher, especially in busy mixes. For sampling from a track in commercial work, there are better options available.

Electronic music with clean, well-defined separation between elements is actually a decent use case for Spleeter. The more separation already exists in the source material, the better any stem splitter performs, and Spleeter is no exception. On a four-on-the-floor track with a clearly defined bass and minimal melodic overlap, the results are fine.

Who it’s still right for

Developers building batch processing pipelines, especially on limited compute, are probably Spleeter’s best current audience. The speed and simple CLI make it easy to integrate, and when you’re processing a large library for analysis, organization, or feature extraction rather than production-quality stems, the quality tradeoff is acceptable.

Researchers who want a reproducible, well-documented baseline for comparisons also have a legitimate reason to use it. Spleeter is cited widely enough in the literature that it functions as a common reference point, and its stability over time makes it reliable for controlled experiments.

Who should look elsewhere

Producers doing sampling, remixing, or any project where the stem quality will affect the final result should use something newer. HTDemucs is the obvious upgrade if you want to stay local and open-source. Online tools are worth considering if managing local model dependencies isn’t something you want to deal with and you want modern output quality without the setup overhead. For a comparison of the free options currently available, this post on free vs paid stem splitters is a useful starting point.

Non-technical users shouldn’t be running Spleeter in 2025. There are too many accessible browser-based options that require no setup and produce better output.

The bottom line

Spleeter deserves credit for what it did for this field. It opened source separation to a wide audience and set a bar that pushed others to do better. The fact that there are now clearly better tools available is partly a consequence of the attention and research Spleeter attracted.

If you’re running Spleeter in 2025 because it was the best option when you set up your workflow, it’s worth revisiting. The landscape has moved significantly, and a few hours of reconfiguring your pipeline could give you meaningfully better results on anything you actually care about.