How to Separate Stems From a Song

Separate a song in four steps

Open the Music Editor and drag your song onto the upload area, or pick a track you already have in your gallery.

Choose Multi-Stem Split from the tool list, then pick a separation type. Vocals + Instrumental returns two files. Full Split breaks the mix into every part the model can find.

Start the job and leave the tab open. Separation runs on the server, so there is nothing to download, install, or configure on your machine.

When it finishes, every stem appears with its own player and its own download button. Take the ones you need and leave the rest.

What you get back depends on where the song came from

This is the part most guides leave out, and it changes the answer. AetherWave runs two different separation engines, and which one handles your song is decided by where the song came from.

A song you upload is separated by Demucs, the open source separation model from Meta AI Research. It works on any audio file you own. A full split returns four stems: vocals, drums, bass, and one combined track carrying guitars, keys, and everything else.

A song you generated on AetherWave is separated by the system that produced it, which can address the parts the track was built from instead of inferring them from a finished waveform. A full split there returns up to twelve stems, with guitar, keys, and strings arriving as separate files.

So the twelve-stem split is not a setting you can switch on for an outside recording. If you are bringing in your own file, four stems is what a full split gives you. If you want a track you can take apart completely later, generate it in Music Studio first.

Formats, file size, and how long a job takes

The uploader accepts MP3, WAV, M4A, FLAC, and OGG, up to 500 MB. Anything that is not already a WAV is converted to one before separation runs, so a compressed source is fine. It just cannot come out better than what the compression left behind.

Separation runs on CPU rather than a GPU, so it is not instant. A full-length track takes several minutes. Jobs are capped at 30 minutes, which comfortably covers a track of about twelve minutes. Anything longer should be cut into sections and run one section at a time.

If a job stops and reports that it ran out of memory, the track was too long to hold in one pass. Trim it to the section you actually need and run that instead. This is the single most common failure and it is fixed by shortening the input, not by retrying the same file.

Stems arrive as MP3, and that matters if you are mixing

Every stem is written out as an MP3. That is a deliberate tradeoff: MP3 stems preview and transfer far faster, and for most of what people do with stems, sampling, a remix sketch, a karaoke bed, a part to practice against, the difference is inaudible.

It does matter if you are pulling stems into a DAW to mix something you plan to release. Separation artifacts and compression artifacts stack, and the vocal stem is where you will hear the result first.

A WAV version of a stem is available on the Studio plan and above, but be clear about what it is. The MP3 is transcoded up to WAV; it is not a lossless original being unlocked. The separation itself wrote MP3, so there is no higher-fidelity file hiding behind it. See pricing for which plans include WAV downloads.

What separates cleanly and what fights back

Clean results: a modern pop, rock, hip-hop, or R&B mix with a distinct lead vocal and parts that occupy different frequency ranges. This is the material the model was trained on and it is where separation looks like magic.

Harder: dense walls of guitar, heavily layered backing vocals, and anything drenched in reverb. Reverb is the giveaway. The model can find the vocal but not the room it was sung in, so the tail smears into the instrumental side.

Hardest: live recordings with crowd noise, mono sources, orchestral material with dozens of parts, and vocals processed so heavily that they read as an instrument. Expect audible bleed rather than a clean lift.

One practical note: judge the result on the stem you actually need. An instrumental hides a great deal because the artifacts sit where the vocal used to be and the rest of the mix masks them. An isolated vocal exposes everything, because there is nothing left to hide behind.

Explore

Frequently Asked Questions

Can I separate stems from any song?

You can separate any audio file you upload, in MP3, WAV, M4A, FLAC, or OGG, up to 500 MB. Quality of the result depends on the mix, not the format. Separation gives you the audio; it does not give you rights to a recording you do not own, so use your own material or material you have permission to use.

How many stems do I get?

A song you upload returns four stems on a full split: vocals, drums, bass, and everything else combined. A song generated on AetherWave can return up to twelve, including guitar, keys, and strings as separate files. Both paths also offer a two-stem split of vocals and instrumental.

How much does stem separation cost?

Stem separation costs 15 credits per job and is free on the Artist plan and above. The current rate is listed on the AetherWave pricing page.

Do I need to install any software?

No. Separation runs on AetherWave servers and the stems appear in your browser when the job finishes. There is no plugin, no DAW, and no local model download.

Are the stems lossless?

No. Stems are written as MP3. A WAV version can be downloaded on the Studio plan and above, but that WAV is transcoded from the MP3 rather than recovered from a lossless original, because the separation step itself produced MP3.

How long does stem separation take?

A full-length track takes several minutes, since separation runs on CPU. Jobs are capped at 30 minutes, which covers tracks up to roughly twelve minutes. Longer pieces should be split into sections first.