Tutorials Video

Separating Stems with AI: Vocals, Drums, Bass and the Rights Question

By Infinite Creation · Intermediate · ~13 min

A 1960s mid-century comic book style editorial illustration showing a music producer at a console as a single record splits into four coloured ribbons of sound that run onto separate tape reels, while a guitarist looks on.

Overview

Stem separation splits a finished song into parts such as vocals, drums and bass. A stem is one of those parts saved as its own audio file. The tools are used by remixers, musicians who want to practise along, and editors who need dialogue apart from music. This guide is for anyone who has a finished mix and no original tracks. It explains which tools state which stems, what free routes exist, what hardware one popular free tool needs, and the rights question you should answer before sharing results. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, and the Sources section says what each page established.

What You Need

A finished audio file you own or have permission to process, in the best quality you hold
For the free desktop route, a computer with the graphics card the maker lists
A way to listen to each stem alone, and then all stems together
A decision about what you will do with the stems: private practice, a remix, or publication

Steps

Answer the rights question first

Splitting a track does not change who owns it. A vocal stem from a commercial song is still the song. Private practice is one situation. Publishing a remix, sampling, or uploading stems is another and needs permission from the rights holders. The music licensing page covers the permissions involved.

Pick the route that matches your computer

Ultimate Vocal Remover is a free desktop app whose code is under the MIT licence. Its page states GPU conversions need at least an NVIDIA GTX 1060 with 6 GB of memory, with 8 GB recommended, and notes AMD support is limited. It runs on 64-bit Windows 10 or later, or macOS Big Sur or later. If your computer does not meet that, a web service or an editor with a built-in feature is the alternative.

Know what stems each option offers

Moises states it separates into up to 27 stem types, including vocals, guitar, bass, drums, piano, keys and winds, with apps for browser, desktop and phones. LALAL.AI lists vocal and instrumental, drums, bass, electric and acoustic guitar, piano, synthesizer, strings and winds. Audacity's OpenVINO effects split a song into vocals and instrumental, or into vocals, drums, bass and other. AudioShake, aimed at film and music businesses, lists dialogue, music and effects separation.

Understand the free plans before you upload

LALAL.AI describes its free Starter plan as mainly for testing, with no full-result download. Moises states a free plan with a limited number of uploads each month. Neither page was tested here. Check what a free plan lets you keep before you build a project on it.

Run it, then listen for bleed

Bleed is a sound from one instrument audible in another stem, such as a faint hi-hat in the vocal. Listen to each stem alone, with the others muted, at a loud but safe level. Listen for a thin, watery vocal and for drums missing their attack. If bleed is audible, try a different model, or a different tool, before editing by hand.

Save with clear names

Name each file with the song, the stem and the tool, for example song_vocals_uvr.wav. Keep the original mix beside them. Do not upload stems of music you do not control.

Pro Tips

Dense mixes with heavy reverb separate worst. A dry, sparse mix gives cleaner stems.
If a desktop tool runs out of memory, the Ultimate Vocal Remover page says lowering the Segment or Window size usually resolves it.
A lossless input file gives the tool more detail than a low-bitrate download.

Knowledge Base

What You'll Learn

The result depends on how the model was trained, and on what the original mix did to the sounds.

Why separation is never exact

A mixed song is the sum of its parts, and summing loses information. Two instruments in the same frequency range cannot be pulled apart perfectly, so the model estimates. Reverb makes this worse, because it smears each sound across time. This is why tools describe their results as stems you will still need to clean up, and why bleed is normal.

Music stems and dialogue stems are different markets

Moises and LALAL.AI aim at musicians. AudioShake states it serves film, television and music, including dialogue, music and effects separation for localization and captioning. Its customer logos and accuracy figures are its own claims and were not checked. For a film or broadcast need, dialogue-music-effects stems are the standard delivery. A dubbing house needs the music and effects without the original voices.

Where This Fits

Separated stems arrive without the balance choices of the original mix, which Mixing a Full Song walks through from gain staging to stopping.

FAQ

Q: Is Ultimate Vocal Remover safe and free?
A: Its page says the GUI code is MIT-licensed and asks users to credit UVR. We did not test it or audit it. Download only from the project's own page.

Q: Is a remix made from separated stems allowed?
A: Only with permission from whoever owns the recording and the song. Separation does not change ownership.

Q: Which tool gives the cleanest stems?
A: We did not test any, so we cannot say. Try your own file in two tools and listen for bleed.

Vocabulary

Stem One group of related sounds from a mix, saved as its own audio file, such as all the drums or the lead vocal. Mixing engineers deliver stems so others remix, master or localize a project.

Bleed Sound from one source audible in a stem meant to hold another. In separated stems it appears as a faint copy of the missing instrument. Some bleed is normal because the separation is an estimate.

Dialogue, music and effects stems The three-way split of a film soundtrack used for foreign-language versions. A dubbing studio replaces the dialogue stem and keeps the music and effects stems from the original.

Sources

Each maker's own page was read on 8 October 2026. We tested no tool. Checked 8 October 2026.

Ultimate Vocal Remover, project page. States MIT licence for the GUI code, minimum NVIDIA GTX 1060 6 GB for GPU conversions with 8 GB recommended, limited AMD support, 64-bit Windows 10 or macOS Big Sur or later, and the memory fix.
Moises. States up to 27 stem types, browser, desktop and mobile apps, and a free plan with a limited number of uploads.
LALAL.AI. Lists the stems offered, a local VST plug-in, and a free plan for testing with no full-result download.
Audacity, OpenVINO AI effects. Lists music separation into vocals and instrumental, or four stems, running locally on a Windows download.
AudioShake. Lists dialogue, music and effects separation, music stems, a browser Studio, an API and a real-time SDK.

Steinberg's SpectraLayers page returned no product text and is not covered here. No page we read states audio quality results. The rights advice is general guidance and is not legal advice.