How to Split a Song Into Stems

A finished song is a cake. Stems are the ingredients — vocals, drums, bass and everything else — separated back out so you can mute the singer for karaoke night, solo the bassline to learn it, or lift a vocal for a remix. AI made this possible from an ordinary MP3 or WAV, no studio files required. Here is how it works and how to do it free, in your browser.

What is a stem, exactly?

In the studio, a stem is a grouped sub-mix exported as its own stereo file: all the drums together, all the vocals together, the bass on its own, and a catch-all "other" for guitars, keys, synths and strings. Producers trade stems so a remixer can rebuild a track without juggling a hundred raw recordings.

Do not confuse stems with multitracks. Multitracks are the original session recordings — every microphone, every take, sometimes 50+ tracks — and they almost never leave the studio. When you separate a released song, AI produces estimated stems from the finished stereo mix. They are remarkably good, but they are reconstructions, not the studio's files.

Stems vs splitting audio into parts

These two phrases get mixed up constantly, so let us be precise. Splitting audio into parts cuts one file into sequential pieces — chapters, tracks, segments along the timeline. Stem splitting separates layers that play at the same time — the voice comes off the top of the drums, for the whole song at once. Cutting a cake into slices vs un-baking it into eggs, flour and sugar. If you want chapters, use the audio editor's Split tab. If you want the singer gone, read on.

What people actually do with stems

• Karaoke & instrumentals: remove the vocal and sing over the real backing track instead of a cheap MIDI remake.
• Remixes, mashups & sampling: lift a clean-ish vocal (an acapella) or isolate a drum break to build on.
• Practice & play-along: mute your own instrument and play its part yourself, at the real tempo, with the real band.
• Learning & transcription: solo the bass or the "other" stem to hear lines that are buried in the full mix.
• DJ edits: strip a track down to drums for transitions, or extend an intro with its own percussion.
• Content creation: keep a song's energy under your video with the vocal removed, so it does not fight your narration. (Check the rights for anything you publish — separation does not change who owns the song.)

How AI stem separation works

Modern separators like HTDemucs — the model our stem splitter runs — are neural networks trained on thousands of songs where the true stems were known. Shown a new mix, the model predicts each stem's waveform directly. It "knows" what voices, kick drums and basslines typically look like in a spectrogram, and it rebuilds each layer sample by sample.

Two honest consequences. First, bleed is normal: a faint ghost of the drums in the vocal stem, especially where reverb smears everything together. Second, the model can only work with what survives in your file — a 96 kbps MP3 has already thrown away detail the AI could have used. Neither is a deal-breaker for karaoke, practice or remixing; both are worth knowing before you judge the result.

How to split a song into stems (step by step)

Using the free AI stem splitter — everything runs on your device, nothing uploads:

1. Load the song. Drop an MP3, WAV, M4A, OGG or FLAC onto the page (up to 12 minutes). The file is decoded locally and resampled to 44.1 kHz, the model's native rate.
2. Wait for the model — once. On your first visit the AI model (~172 MB) downloads and is cached in your browser. Later visits skip straight to work.
3. Let it separate. The song is processed in short segments; a progress bar counts them off. A typical song takes under a minute on a desktop, longer on a phone.
4. Audition in the mixer. Four stem cards appear — vocals, drums, bass, other — each with its own waveform, mute, solo and volume. Solo the vocal: that is what the AI heard as "the singer".
5. Download what you need. Every stem as WAV (lossless) or MP3, an Instrumental button (drums + bass + other — your karaoke track), or all stems at once. On phones you can also share a stem straight into another app.

Five tips for cleaner stems

• Feed it lossless when you can. Same song, better input, better separation — this is the single biggest quality lever you control.
• Avoid clipped, loudness-war masters. Heavily limited mixes glue instruments together; the model struggles to pull them apart. A quieter, more dynamic master separates better.
• Expect "other" to be busy. Guitars, keys, synths, strings — and usually the backing vocals' reverb — live there. It is a catch-all by design.
• Live recordings bleed more. Room reverb and microphone spill are baked into every layer. Studio recordings separate cleanest.
• Check loudness before you publish a stem. A solo stem is usually quieter than the full mix. Run it through the loudness analyzer and know its integrated LUFS before it goes anywhere near a platform target like Spotify's -14 LUFS.

Frequently asked questions

Is the stem splitter really free?

Yes. The AI model runs entirely in your browser, so there is no server bill to pass on to you: no account, no minutes, no per-song credits. Your song is never uploaded — the separation happens on your own device.

Why can I still hear faint drums in the vocal stem?

That is bleed, and it is normal for AI separation. The model estimates what belongs to each stem from a finished stereo mix; reverb, effects and instruments that overlap the voice in frequency leave faint traces. Starting from a high-quality source (WAV or FLAC rather than a low-bitrate MP3) reduces it noticeably.

How do I make a karaoke track from a song?

Separate the song, then use the Instrumental download — it combines the drums, bass and other stems, which is the original mix minus the vocals. In the built-in mixer you can also mute just the vocals and listen before downloading.

What is the difference between stems and multitracks?

Multitracks are the studio's original recordings — every microphone and instrument on its own track, sometimes 50 or more. Stems are grouped sub-mixes (all drums together, all vocals together). AI separation produces estimated stems from a finished mix; it cannot recover true multitracks that were never released.