How to Split a Song into Stems Online: Vocals, Bass, Drums
Need an instrumental for a live show, an acapella for a remix, or an isolated bass line to learn a part? All of that can be made from an ordinary finished mix — no source files, no DAW. Here is how to split a song into stems online in Magic Master and what to honestly expect.
What stems are and what you get
A stem is an isolated layer of the mix. Our separation returns four:
- Vocals — all vocals including backing;
- Drums — drums and percussion;
- Bass — the bass line;
- Other — everything else: guitars, keys, synths, strings.
The sum of the four stems is practically equal to the original track — mute and mix them in any combination. Vocals + Other without Drums/Bass gives a "bed", Drums + Bass + Other is a classic instrumental, Vocals alone is an acapella. You can then bring the acapella to an even level with the Dry vocal preset — that is vocal mastering online, with no reverb and no added colouration.
How it works under the hood
The separation runs on Demucs (Meta AI research) — a network trained on thousands of tracks with known sources. It "hears" a mix roughly the way an engineer does: it recognizes the timbre of a voice, the attack of a kick, the character of a bass. Processing happens on our server; your file is not shared with third parties or published.
In our measurements the sum of stems reconstructs the original at about 28 dB SNR — a strong result for neural separation of a finished mix.
Step by step
- Open the app and upload a track (WAV, FLAC, MP3 — up to 200 MB).
- In the PRO modules block press "Split into stems · 1 token".
- Wait for processing — about 4× the track length (a 4-minute song takes around 17 minutes). If the server is busy, you see your queue position and start estimate. You don't have to stay on the page: the archive is kept in your profile for 24 hours.
- The archive downloads automatically; with Telegram linked it also arrives in your chat.
The token is charged only after successful separation — a failure costs nothing.
Honest about quality
Neural separation is not the original studio multitrack:
- dense mixes (wall of sound, heavy compression) may show light "musical noise" in quiet spots of stems;
- reverb "glued" to the vocal at mixdown will partially stay in the vocal stem;
- a full rebuild of a track needs source files — stems are for instrumentals, remixes, sampling and learning.
More on artifacts — in AI separation artifacts explained.
What to do with stems next
- Instrumental/karaoke — mute vocals, mix down the rest.
- Remix — take the vocal or the rhythm section as material.
- Stem mastering — process the layers separately: we also offer stem mastering.
- Arrangement study — listen to parts in isolation.
Try it on your track: magicmaster.pro/stems-separation.
Questions and answers
Can vocals be extracted from any song?
From any finished mix — yes: the Demucs neural network splits a track into vocals, drums, bass and other. Quality depends on arrangement density: typical pop/rock/electronic comes out close to studio quality, while a dense wall of sound may show light 'musical noise' in quiet spots.
How much does it cost and how long does it take?
1 token per separation, charged only after success. Server processing takes about 4× the track length: a 4-minute song — around 17 minutes. The screen shows the time left, and you can close the tab: the finished archive stays in your profile for 24 hours.
What do I get exactly?
A zip archive with four WAVs: vocals, drums, bass, other. The link is kept for 24 hours; with Telegram linked, the archive also arrives in your chat.
Is it legal to make an instrumental from someone else's song?
For personal use (karaoke, learning parts) — yes. Publishing remixes or covers of copyrighted material requires the rights holder's permission — that responsibility is yours.
Upload a track — a finished master in about a minute.
Open mastering →AI mastering onlineLUFS analyzer