Split a Song into Stems: Vocals, Bass, Drums

Separate any song into vocals, bass, drums, and everything else — then download each stem individually for karaoke, practice, or remixing.

A "stem" is one isolated layer of a song: the vocals alone, the drums alone, the bass alone. Studios have these from the original session. The rest of us have a finished MP3 and a dream. The Stem Splitter closes that gap, it pulls any song apart into four stems right in your browser: vocals, bass, drums, and other (everything else).

Each stem gets its own preview and download. Instant karaoke backing track, an isolated drum groove to practice with, a vocal for your remix, all from one upload. One thing to know going in: this is classical DSP signal processing, not AI. Expect some bleed between stems, especially on dense mixes. More on that below, honestly.

Open Stem Splitter →

The four buckets

Upload a song, get four downloadable parts:

Vocals, the lead voice, isolated as cleanly as math allows. Mute it for karaoke, keep it for the remix acapella.
Bass, the low end: bass guitar, synth bass, the thump of the kick.
Drums, the kit: snare crack, hats, cymbals, kick attack.
Other, the rest of the band: guitars, keys, strings, pads, effects.

The split is energy-conserving, which is a fancy way of saying the four stems add back up to the original mix exactly. Nothing lost, nothing double-counted. Mute the vocal stem and what’s left is a true instrumental, that’s the whole karaoke trick.

Splitting a song

Open the Stem Splitter and upload your song. MP3 or WAV, a clean stereo master works best.
Run the split. Spectral processing does the surgery. It takes longer than most tools on this site, so give it a minute. Watch the progress bar, not the clock.
Preview every stem before downloading anything. Ears first, always.
Download what you need. Each row has its own button, the drums alone, the instrumental, all four, whatever the project wants.

After karaoke specifically? Download everything except the vocals, then glue the remaining three together with the Audio Joiner. Or just play the three stems at once, same result, less work.

The DSP-vs-AI talk

Let’s be straight about this. The flashy AI splitters use neural networks trained on thousands of multitrack studio recordings. This tool uses classical DSP, spectral gating, frequency masking, the old-school stuff, running entirely in your browser. No server, no model download, no queue, no bill.

The price is bleed. You’ll hear ghostly traces of other instruments hanging around in each stem: snare rattling faintly in the vocal, a breathy vocal smearing into "other." Clean, well-mixed stereo masters separate best; dense, heavily effected, or mono recordings separate worst. For karaoke night, practice sessions, and sketching remixes, it’s genuinely useful. For a studio-grade acapella you’d release? A dedicated AI service earns its money there. Horses for courses.

Getting cleaner splits

Feed it quality. A 320 kbps MP3 or WAV separates noticeably better than a 128 kbps stream rip. Compression artifacts look like musical content to spectral analysis, and the tool can’t tell the difference. Garbage in, ghosts out.
Center-panned vocals win. The vocal detector leans hard on the fact that lead vocals usually sit dead center. Hard-panned backing vocals will wander off into "other", that’s physics, not a bug.
Sparse beats dense. Voice plus acoustic guitar pulls apart beautifully. A wall-of-sound metal mix? Everybody’s fighting over the same frequencies. Manage expectations accordingly.
Audition before committing. If the vocal stem’s drum bleed ruins your remix idea, try a different section of the song, verses are usually cleaner than choruses, or find a better source file.

Beyond karaoke

Karaoke is the gateway drug. The rest of the pharmacy:

Practice parts. Learning bass? Solo the bass and drums and play along, it’s the band in your room, minus the scheduling. Drummers: loop the drum stem and study fills bar by bar.
Sample fodder. Chop drum breaks and basslines from isolated stems instead of wrestling the full mix. Clear the rights before you release anything, "but the splitter gave it to me" impresses no lawyer.
Mashups. Vocals from song A over the instrumental of song B, after checking both with the BPM & Key Finder. Compatible key first, arrangement second.
Vocal chops. Slice one word from the vocal stem, repeat it into a hook. Half of modern electronic music is this trick.
Live stem blending. Some DJ software takes stems natively: drop track A’s drums out while track B’s vocals come in. The crowd hears magic; you hear four faders.

Even imperfect stems are plenty for sketching, and every finished track started as a sketch somebody didn’t delete.

Frequently asked questions

Is this as good as AI stem separation?

No, and I’d rather tell you than have you find out mid-project. AI models trained on multitrack recordings separate more cleanly. What you get here instead: instant, private, free, in your browser, no uploads, no queues. The trade is audible bleed on complex mixes. Pick the right tool for the job’s standards.

What exactly are the four stems?

Vocals (lead voice), bass (low-end instruments), drums (kit and percussion), and "other" (guitars, keys, strings, effects, the rest of the band). Energy-conserving, so the four sum back to the original mix exactly.

Can I make a karaoke track with this?

Yes, it’s the most popular use. Split the song, download bass, drums, and "other," combine them with the Audio Joiner, and you’ve got an instrumental. On clean pop productions it works surprisingly well; on dense mixes expect faint vocal ghosts. Preview first.

Why is there drums in my vocal stem?

Bleed. Snare hits and cymbals live in the same frequency neighborhood as vocal sibilance, so spectral separation can’t fully untangle them. It’s worst on loud dense choruses, mildest on sparse verses. A better source file is the only real fix.

Does my song get uploaded to a server?

Nothing is uploaded, ever. The separation runs locally in your browser, that privacy is precisely why the tool uses DSP instead of phoning a server-side AI model. Your music stays your music.