Removing the vocal from a finished song used to mean a trick with phase cancellation that only half worked. Today an AI model can listen to a mixed track and rebuild it as separate layers, called stems: one with the voice and one with everything else, or one per instrument.
This guide explains what actually happens when you separate a song, how to get the cleanest result, and what the limits are. The steps use the stem splitter in this studio, but the advice on source files and clean-up applies to any tool.
What vocal removal really does
A mixed song is a single waveform: the voice, drums, bass and guitars are summed together. There is no hidden vocal track inside an MP3 that a tool can simply switch off. A separation model has learned, from a very large amount of music, what a voice sounds like compared with a snare, a bass line or a pad. It predicts which parts of the sound belong to which source and writes each one out as its own file.
Because it is a prediction, the result is very good but not perfect. You will usually get a clean, usable vocal and an instrumental that sounds like the original minus the singer. In dense passages a little of one part can remain in another, for example a cymbal shimmer in the vocal stem. This is called bleed, and there are simple ways to deal with it (see the last section).
The old phase-cancellation trick (inverting one stereo channel against the other) only removes what is panned dead centre, and it takes the kick, bass and snare with it. AI separation keeps them.
Before you start: rights and the source file
Only separate audio you own or have permission to use: a song you generated here, your own recording, or a track licensed to you for this purpose. Splitting a commercial release you downloaded doesn't make the parts yours, and every upload in the studio asks you to confirm your rights before it is used.
The quality of the file you start from matters more than any setting. Separation works on what is actually in the audio, so missing detail cannot come back.
- Use the original export when you have it: WAV or FLAC, or an MP3 at 256–320 kbps.
- Avoid copies recorded from speakers or ripped from a video at low quality; the noise and compression end up in every stem.
- Don't add effects before separating. Normalising is fine; heavy limiting or reverb makes the parts harder to tell apart.
- Accepted formats: MP3, WAV, M4A, MP4, AAC, OGG, FLAC, WEBM, up to 20 MB per file. Uploads are deleted from storage after seven days; the stems stay in your library.
Two stems or a full split?
Decide what you need the parts for before you choose a mode, because the two modes cost different amounts and give different files.
- Two stems (one credit): a vocal file and an instrumental file. This is all you need for karaoke, an acapella for a remix, or a backing track to sing over.
- Multi-stem (four credits): up to 12 separate parts, such as lead vocals, backing vocals, drums, bass, guitar, keys, percussion, strings, synth, brass, woodwinds and effects. Use it when you want to practise along with one instrument, rebalance a mix, or sample a single part.
Multi-stem only returns parts that are really present. A piano-and-voice ballad will not come back with a drum file, so the number of stems you receive depends on the arrangement.
Step by step
- Open the stem splitter. You can browse it without an account; separating needs a paid plan or any past credit purchase.
- Pick a song from your library, or upload a file you have the rights to and tick the confirmation box.
- Choose two stems or multi-stem. The credit price is shown before you confirm.
- Wait for the job to finish; it usually takes a few minutes at most. If it fails, the credits return to your balance automatically.
- Play each stem in the browser, then download the files you want and bring them into your editor or DAW.
What to do with the stems
- Karaoke: play the instrumental stem and sing over it, or turn it into a practice track by lowering the tempo in your editor.
- Remix: drop the vocal stem onto a new beat. Match the tempo first, then the key, so the voice doesn't have to be stretched far.
- Practice: solo the bass or guitar stem to hear exactly what is played, then mute it and play the part yourself.
- Mix fixes: if the vocal of a song you generated sits too low, separate it and rebalance the two stems.
- New arrangements: keep the vocal and use add instrumental to put different backing music under it.
Cleaning up bleed and artefacts
Most stems are ready to use as they are. When something does leak through, a few light touches in any editor usually solve it:
- Cymbals or hi-hats in the vocal: a gentle high-shelf cut above about 10 kHz, or a de-esser set lightly.
- A faint voice left in the instrumental: it is usually masked once you add your own vocal or another layer on top. If not, a narrow cut around the singer's strongest frequencies helps.
- Watery or phasey sound: this comes from a low-quality source. Re-run the split from a better file rather than trying to repair it.
- Silence between phrases in the vocal stem: a noise gate keeps those gaps clean.
Remember that the stems carry the same license as the song they came from. For songs made here that is personal, non-commercial use; the Terms of Service have the details.



