Friday night, living room, and the only thing missing is the version of the song without the singer in it. Making a karaoke track used to mean hunting for an official instrumental that usually doesn't exist — or muting one channel of a stereo file and hoping. AI stem separation replaced both hacks, and it now runs in a browser tab without the song going anywhere near a server. Here's the honest version: what separation can and can't do, which engine fits your night, and the whole flow in four steps.

Split a song in your browser →

AI vocal removal that runs locally — instrumental and vocal in one pass. Free daily quota, no sign-up to start.

Three jobs, three quality bars

"I need the vocals gone" is three very different jobs, and each judges the result differently:

  • The host — Friday night, people singing over the speakers. The instrumental only has to hold up behind live voices in a room, which is exactly where separation artifacts are least audible. Speed and plays-on-anything beat perfection here
  • The cover artist — you want the original vocal out so yours can go in, or isolated to study how the singer phrases the bridge. Your judge is a pair of headphones up close, and artifacts that vanish at a party get exposed here
  • The producer — stems for a remix or a sample flip, the least forgiving audience of the three: every ghost vowel left in the instrumental is one you'll EQ around for the rest of the project. There's a dedicated acapella route for exactly this crowd

What separation actually does — manage the expectation once

No tool reaches into a finished mix and lifts the microphone off the singer; the recording is already blended. What the model does is estimate, moment by moment, which energy belongs to the voice and which belongs to everything else — then hand you two reconstructions. On a clean mix the estimate is startlingly good. On a wall-of-sound production with drenched reverb, the edges blur: faint vocal residue can haunt the instrumental, and a little cymbal shimmer can travel with the vocal. Knowing this beforehand is the difference between "the tool failed" and "that mix is genuinely hard." Why the two engines differ comes down to model size and hardware.

The whole job in one browser tab

  1. Open the splitter and drop the song in — an audio file, or the audio pulled straight out of a video (that trick, explained)
  2. Pick an engine — the fast default runs on any CPU including phones; the heavier one wants a WebGPU browser (Chrome or Edge) and a shorter song
  3. Split. Both stems arrive in a single pass — instrumental for the party, vocal for everybody else
  4. Download, then listen once on the equipment the room will actually use — before the room fills up

The split runs free on a daily quota per tool; signing in raises it. The song never leaves the machine — for unreleased masters and not-yet-announced covers, that's the only acceptable arrangement anyway.

Choosing the engine in one line each

  • Default engine (Spleeter) — the host's choice and the cover artist's first draft: runs on any CPU including phones, takes songs up to ~30 minutes on desktop, and at a karaoke night the gap to the premium engine disappears behind the singers
  • Heavy engine (Demucs, 4-stem) — the producer's pick and the headphone-auditioning artist's: visibly cleaner on dense mixes and reverb-heavy vocals, songs up to ~8 minutes, wants a WebGPU-capable browser
  • Phones — fine for a quick split, but songs cap at 3 minutes and the heavy engine is off the table. Real karaoke prep happens on a laptop

Finishing touches that make it play like a real karaoke track

  • Trim before you split (either order works). The 12-second silent intro and the long outro tail are dead weight at a party — cut the segment first so the file getting passed around starts where the singing starts
  • Mind the loudness gap. An instrumental inherits the original mix's balance and can sit a touch quieter than commercial karaoke tracks; a quick level check before guests arrive beats apologies mid-chorus
  • The key stays the key. Separation changes what's in the file, not where it sits — if your singer needs the song lower, that's a pitch-shifting player feature, not something splitting should ever do to the audio

Quick answers

  • Is it really free? The split runs on a free daily quota per tool — enough for a night's setlist — with no sign-up to start; signing in raises the cap
  • Is my song uploaded? No. The model loads into the browser and runs on the device; the audio never leaves it
  • How long can the song be? Around 30 minutes per song on desktop with the default engine, about 8 with the 4-stem one; phones cap at 3 minutes
  • Do I get both stems in one pass? Yes — one run produces the instrumental and the vocal together
  • Can I perform or post the instrumental? Private parties, practice and remix sketches are what these tools are for. Public performance or monetizing the stems steps into licensing territory — separating them doesn't change who owns the song

Split a song in your browser →

AI vocal removal that runs locally. Free daily quota, no sign-up to start.