What Are Audio Stems?

A stem is a single, self-contained track that holds one part of a song — all the vocals, or all the drums, or all the bass — kept separate from the rest of the mix so it can be handled on its own. A finished song you hear on the radio is one blended stereo file, but before it was blended it existed as a stack of these separate parts. Stems are those parts: the ingredients, bottled individually, instead of the finished dish.

Understanding stems is the key to a huge amount of music work — remixing, karaoke, sampling, spatial audio, live performance. But the word gets used loosely, and it means slightly different things depending on whether you're recording a song or taking one apart after the fact. Let's untangle it.

Where the word "stem" comes from

The term is older than any software you're using. It goes back to multitrack tape, when a recording studio captured each performance — the singer, the drum kit, the bass, the guitars — onto its own separate track. During the final mixdown, an engineer balanced all those tracks together into the finished stereo master.

The problem was that mixing dozens of individual tracks at once, every time, was unwieldy. So engineers started bouncing groups of related tracks down to a smaller set of submixes — all the drum microphones summed to one drum track, all the backing vocals to one vocal track, and so on. Those grouped submixes were the original "stems": intermediate bundles, sitting partway between the raw multitrack and the final two-channel master. In film post-production the same idea survives almost unchanged — a movie's audio is delivered as a dialogue stem, a music stem, and an effects stem, so it can be re-balanced or re-dubbed into other languages without remixing from scratch.

The modern music meaning has drifted a little. Today, when most people say "stems," they mean the separated instrument-group tracks of a song — and increasingly they mean tracks pulled out of a finished mix rather than bundled together while making one.

The four standard stems

There's no law that fixes how a song must be divided, but one split has become a de facto standard, largely because it's what AI separation tools produce: vocals, drums, bass, and "other."

• Vocals — every sung or spoken part, lead and backing, including the reverb and effects printed onto them.
• Drums — the whole percussion kit: kick, snare, hats, cymbals, and usually any other rhythmic percussion.
• Bass — the low-end instrument holding down the root of the harmony, whether that's an electric bass, a synth bass, or an upright.
• Other — everything left over. Guitars, pianos, synths, strings, horns — all the melodic and harmonic material that isn't voice, drums, or bass gets pooled here.

That last "other" bucket is a giveaway that this four-way split is a practical compromise, not a musical truth. A real arrangement might have twenty distinct instruments; grouping the non-rhythm-section melodic parts into one catch-all is simply the division that turned out to be both useful and reliably achievable. A simpler two-way version — just vocal and instrumental — is the most common request of all, and it's really the four-stem split with drums, bass, and other summed back together.

Stems you record vs. stems you extract

This is the distinction that clears up most of the confusion. There are two completely different ways to end up with stems.

Stems you record. If you make the song yourself in a DAW, the stems already exist — every track you laid down is a stem, or can be bounced into one. Exporting them is lossless and perfect, because the parts were never mixed together in the first place. This is how remix contests and official remix packs work: the original artist hands over the true, clean stems straight from the session.

Stems you extract. If all you have is a finished MP3 — someone else's song, a track whose session files are long gone — the parts were baked into one file years ago, and getting them back means un-mixing the audio. That's a fundamentally harder, estimation-based job done by AI source separation, and the results are astonishingly good but never bit-perfect the way recorded stems are. If you want the details of how a neural network pulls that off, how AI vocal removal actually works walks through the mechanism — and what a spectrogram is explains the frequency picture those models actually operate on. The short version: the model has learned what a voice, a drum, and a bassline each sound like, and it uses that knowledge to sort every moment of the audio into the right bucket.

The practical upshot is that extracted stems can carry faint traces of each other — a whisper of cymbal in the vocal, a little smear on a sustained synth. Recorded stems never do. It's worth knowing which kind you have before you build something on top of them.

What people actually use stems for

Stems are one of those quiet foundations that a surprising number of things are built on:

• Remixing and mashups. Rebuilding a track means getting at its parts. A remixer might keep only the vocal and write an entirely new instrumental underneath it, or layer one song's acapella over another's beat. Pulling a clean vocal is the classic starting point — how to make an acapella from any track covers exactly that.
• Karaoke and backing tracks. Drop the vocal stem, keep everything else, and you have an instrumental to sing over. That's the entire basis of karaoke, and a live karaoke mixer that balances vocal against instrumental in real time leans directly on the split.
• Sampling and beat-making. Producers isolate a single drum groove or a bassline to chop, loop, and reuse. Splitting a track into four stems hands you the drums alone — with no melody bleeding in — which makes the sample far cleaner to work with.
• Spatial and immersive audio. Formats like Dolby Atmos place individual elements around a 3D field. You can't position a voice behind a listener if it's welded to the guitars; immersive mixes are built from stems by design.
• Live performance. Bands play to backing tracks that supply the parts nobody on stage is covering — strings, extra synths, doubled vocals — while muting the parts the musicians play live. That selective playback is only possible because the parts are separate stems.
• Cleanup and restoration. Separating a song lets you treat one part without touching the others. You can denoise a vocal in isolation, the same way AI noise reduction targets voice, then fold it back into the mix untouched.

Getting stems from a song you didn't record

If you're working from a finished track, an in-browser separator is the fastest way in. A four-stem splitter that separates vocals, drums, bass, and other hands you all four parts to mix and mute independently; if you only need the two-way divide, a vocal remover that outputs a clean vocal and instrumental gives you that directly, ready to sing over in the live karaoke mixer.

On vocalcut.com both run on one HTDemucs model through onnxruntime-web, entirely inside your browser tab. The first time you use it, the model itself downloads to your device (tens of megabytes) and then runs on your own CPU — so the very first separation isn't instant, though later ones skip the download. The trade worth the wait: your audio never leaves your device. Nothing gets uploaded to a server, because the un-mixing happens locally. What downloads is the tool coming to your file, not your file going to the tool.

Frequently asked questions

Is a stem the same as an individual track? Not quite, though people use the words interchangeably. A track is usually one recorded source — one microphone, one instrument. A stem is a group of related tracks bounced together: the drum stem might contain a dozen individual drum-mic tracks summed into one. In everyday use, especially with AI-separated audio, "stem" just means one of the separated parts of a song (vocals, drums, bass, or other), so the distinction rarely matters outside a studio.

How many stems does a song have? There's no fixed number — it depends on how you divide it. The recording engineer chooses the groupings, and AI separation tools settle on a standard four: vocals, drums, bass, and other. A simpler split gives you just two (vocal and instrumental), while a full multitrack session could be broken into dozens. Four is the practical default because it's both useful and reliably achievable.

Are extracted stems as good as the original recorded stems? No, and it helps to expect that. Stems exported from the original DAW session are perfect, because those parts were never mixed together. Stems pulled out of a finished file are estimated by an AI model un-mixing the audio, so they can carry faint traces of neighbouring parts — a hint of cymbal in the vocal, a little texture on sustained notes. They're remarkably clean and good enough for the vast majority of uses, but they aren't bit-for-bit identical to true source stems.

Do I need the original session files to get stems? No. If you have the session files, exporting stems is trivial and lossless. If you don't — which is the usual case for any song you didn't make — an AI separator reconstructs the parts from the finished mix. You upload nothing to a server with a browser-based tool; the separation runs locally on your device and estimates the stems from the audio you already have.