Demucs, Spleeter, and cloud separators: a short, honest history

May 2026 · 8 min read

Stem separation had a before-and-after moment around 2018–2019. Open-source models made it possible to split vocals from instrumentals without hand-tuned phase tricks. Then every web app slapped "AI" on a landing page and stopped explaining which AI.

Spleeter: fast, famous, dated

Deezer's Spleeter was one of the first widely deployed separators — lightweight, fast, good enough for demos and karaoke. It predicts four stems (vocals, drums, bass, other) or two (vocals / accompaniment). Speed and predictable GPU use made it the default for hosted APIs.

It is not the state of the art in 2026. Dense mixes, reverbed vocals, and complex stereo imaging expose its age. Still usable. Still deployed everywhere because infra teams can cost it.

Demucs: heavier, often cleaner

Meta's Demucs line (v3, v4, hybrid transformers) generally wins blind listening tests against Spleeter on challenging material. You pay in compute: longer runs, more VRAM, higher per-minute cloud bills. Offline enthusiasts swear by Demucs for batch work on a local GPU.

"Demucs is better" is directionally true and practically incomplete. A slow, expensive job that you never rerun because you ran out of patience is not better for you.

What SongRemoveVocals runs

We separate vocals via Spleeter on Replicate. That choice balances cost, latency, and output quality for a web tool priced from $4.99 minute packs with a free daily tier. Demucs might squeeze extra clarity on some tracks; it would also squeeze the economics that let us offer 10 free minutes every day without a subscription.

We would rather be honest about the stack than imply a secret model that invents multitracks from thin air.

Cloud API vs local UVR stacks

Ultimate Vocal Remover and similar local GUIs bundle multiple models — Spleeter, Demucs, MDX variants — and let power users swap weights until something clicks. That flexibility is real. So is maintaining Python environments, CUDA drivers, and batch scripts at 1 a.m.

Cloud separators (us, Moises, LALAL.AI, others) trade control for convenience: upload, wait, download WAV or MP3. Right tool depends on frequency and tolerance for ops work. Our comparison article lays out pricing models without declaring a universal winner.

What no model fixes

Mono or phone-speaker recordings
Heavily compressed social clips
Mixes where vocals and lead guitar share everything

Model upgrades shift the ceiling; they do not remove it. Test your catalog on actual songs instead of benchmark screenshots.