Spleeter vs Demucs: Which AI Stem-Splitter Sounds Better?

Formavide — Tue, 27 Oct 2026 03:01:05 +0000

Most "which is better" answers repeat marketing. We ship both engines in one browser tool, run them on the same songs daily, and can tell you exactly where each wins — and why the difference exists.

Try both engines in your browser →

Spleeter for speed anywhere, Demucs 4-stem on WebGPU. Nothing uploads. Free to try.

Two models, two generations of thinking

Both networks do the same job — pull a song apart into vocals and accompaniment — but they look at the problem differently, and that difference explains everything about their results.

Spleeter (Deezer, 2019) works like a picture editor. It converts the song into a spectrogram — a snapshot of which frequencies sound at which moments — and a compact network paints a mask over it: this energy is voice, that is everything else. Fast and light, because masking a picture is cheap.
Demucs (Meta, evolving through 2022's htdemucs) works more like a musician re-performing the song. It processes the actual audio waveform through a deep encoder-decoder, modeling how each instrument evolves over time rather than judging frequencies frame by frame. Heavier — much heavier — and that's where its quality comes from.

Where you actually hear the difference

On a clean mix with a dry lead vocal, both produce very usable stems — many listeners won't pick a favorite blind. The gap opens under stress:

Dense, loud mixes. Spleeter's mask tends to leave ghostly vocal residue in the instrumental and take some cymbal shimmer with the vocal. Demucs keeps the sides cleaner.
Reverb-heavy vocals. Echo smears the vocal across time — a frame-by-frame mask smears its guess with it. Demucs' longer view of the audio tracks that trail far better.
Fasts, transients, drums. In 4-stem mode, Demucs' drum separation is decisively more musical — hi-hats stay hi-hats instead of turning into artifacts.

Speed, weight, and what it costs to run

Spleeter (2-stem) — vocals + instrumental · runs on any CPU, phones included · fast on any laptop · songs up to ~30 minutes · usable on hard mixes, with some artifacts

Demucs (htdemucs, 4-stem) — vocals, drums, bass, other · wants a GPU (WebGPU in Chrome/Edge) · heavier, GPU does the lifting · songs up to ~8 minutes · clearly cleaner on hard mixes

So which should you use?

The honest answer is a decision tree, not a winner:

Quick job, long file, any device — Spleeter. A 20-minute mix for a background track doesn't need (and can't run) Demucs.
Karaoke night, casual covers — Spleeter at default settings is usually indistinguishable in the room, over speakers, with people singing.
Remix or sampling, vocals front and center — Demucs. Every artifact you don't remove now is one you EQ around later.
Testing if a song "can" be separated at all — Spleeter first for the shape of the result, Demucs for the keepable version.

Why we ship both instead of picking one

Single-engine tools make you fit the job to the model. The splitter we build puts both behind one drop-in, running entirely in your browser — Spleeter on any CPU, Demucs on WebGPU-capable machines — and both stems come out of a single pass either way. Because everything runs locally, you can A/B the engines on the same song as many times as you like, with unreleased material never leaving your disk.

What no stem-splitter can do

Recover a vocal that was mixed into a single bounced file with full-band distortion — separation estimates; it doesn't un-mix physics.
Beat stems delivered by the original studio. For release-grade work, cleared acapellas still win; for creative work, practice and remixing, modern separation is genuinely good.