Stems to remix

A finished song is a bit like a baked cake. The voice, drums, bass and everything else were mixed together into one stereo file, and in principle you can't get the eggs back out. Look at one below: the voice, the drums and the bass are all in there, overlapping in the same picture.

For decades, remixers worked around that. Jamaican dub engineers and disco remixers in the 1970s had the original multitrack tapes; everyone else had to hunt for a bar where the vocal sang alone. Then machine-learning separators arrived, and since Deezer's research team released the open-source Spleeter in 2019, anyone can pull an estimate of the vocals, drums, bass and "everything else" out of a mixed song. Estimates, not originals: they're good enough to build with, and imperfect in ways you'll learn to hear.

This project takes one song you're allowed to use, separates it, and rebuilds a minute of it your way: a new arrangement you can play right after the original and say "listen to what I made of this." Before you separate anything: which moment will be hardest for the model, a voice singing alone, or a voice over a crash cymbal?

Stage 1: Choose a song and an idea

Pick a song you made, or one whose owner says you can remix it. Choose thirty to sixty seconds rather than the whole thing; it's easier to finish something short. Then write one sentence about what your remix will do: "a quiet, stripped-back opening", "the vocal over a completely new beat", "half speed and dreamy".

Checkpoint: a clip and one sentence of intent. Stem separation explains what a separator can and can't recover.

Stage 2: Separate it

Drop your clip into the separator above and run it. Some methods process the song on your own computer and some upload it to a server; if the song is private, pick a local one.

Listen to each stem on its own and mark three moments with timestamps. In the vocal stem, listen between phrases for ghostly music leaking through. At a drum hit, listen for a softened attack; on a long held note, listen for a watery, fluttering texture. These marks show where your stems are strong and where they'll need hiding.

Checkpoint: four exported stems, plus a short list of where each one is weak. A stem that sounds rough alone can still be perfect in the background.

Stage 3: Line them up

Import the four stems into your music software, all starting at exactly the same point. Play them together with the original muted: it should sound almost like the song. If something sounds flanged or doubled, a stem has slipped; line it up again.

Keep the exported stems untouched and work on copies. Mute and level choices you made in the separator don't travel with the exported files, so set your levels again in your own session.

Checkpoint: a session where the stems play in sync. Now you have a song you can take apart.

Stage 4: Change the arrangement

This is the fun part. Start your remix with only one or two stems, say the vocal alone over a new chord pad you've made, and bring in more for the second half. Or keep the vocal and drums and replace the bass with your own line. Change one big thing, not ten; a remix with one clear idea sounds intentional. New parts have to share the song's tempo, and a tempo tool may report double or half the one you feel: Beat and tempo tracking shows how to check its grid by ear before you build on it.

Use the weak spots you marked. If the vocal has music leaking in a gap, don't leave that gap exposed; put your new part there. If the drums sound soft, layer a crisp hit of your own on top. Now the stems' flaws are covered by your choices instead of on show.

Checkpoint: two sections that sound like a remix, not a copy. Form and arrangement helps you shape the contrast.

Stage 5: Make room for the vocal

When you add new parts, they fight the vocal for space. Usually the fix isn't to boost the vocal; it's to carve a little room in the parts around it.

A gentle dip in the drums or your new chords, around where the voice is clearest, opens a pocket for it. Compare at matched level so you hear the space, not the volume. Now the vocal floats on top without being turned up.

Checkpoint: your new parts and the vocal both clear at the same time. Making room with EQ goes deeper.

Stage 6: Throw one word into an echo

Pick one word of the vocal, maybe the last word before your second section, and send only that word into an echo. It's the classic dub move and it instantly sounds like a remix.

Here the whole vocal goes through the echo, so you can hear it blur. In your session, open the echo only for one word. Remember that the echo repeats everything in the stem, including any leaked music, so pick a word where the vocal is clean. Now your remix has a signature moment.

Checkpoint: one echo throw that fills a gap. Dub delay throw has the full recipe.

Stage 7: Finish and play it beside the original

Bounce your remix with its full ending. Then play the original and your remix back to back, at similar loudness, so neither wins by being louder. Ask one question: does the remix do what your sentence from Stage 1 said?

Listen to your marked weak spots one last time. If a flaw still sticks out and bothers you, cover it or cut it. If it adds character, keep it. Then play both versions for a friend.

Checkpoint: a bounce, the original, the stems and your session saved together, so you can come back and make a second version.

Quick check

1. The four stems played together sound almost exactly like the original. Does that mean each stem is correct? No. A bit of vocal can sit in the "other" stem and still add back up perfectly. Listen to each stem alone to judge it.
2. The vocal stem has ghostly piano in the gaps. What are your options? Cover the gaps with your new parts, gate or cut the silences, or use those phrases only in a busy section where the leak is masked. Heavy EQ rarely removes it cleanly.
3. You add an echo to a vocal word and hear a faint echo of the guitar too. Why? Some guitar leaked into the vocal stem, and the echo repeats whatever is in the stem. Choose a cleaner word, or keep the throw short.

Next steps

• Use it: Finishing and references. Compare your remix with the original at matched loudness.
• Sideways: Vocal chain. Give your separated vocal the same treatment a studio vocal gets.