How do you separate vocals from a song?

Import the highest-quality version of the track you have into Veena and run stem separation — the isolated vocal, drums, bass and other instruments land as separate tracks on a real timeline, not as files in a download folder. From there you can build a remix, a cover, a practice track or a sample around them, and export WAV, MP3 or MIDI you own. Free Basic tier; Veena Pro is $20/month.

Why do separated vocals sound watery or metallic?

Because separation models work on a time-frequency picture of the mix, and reverb tails, cymbals and heavily processed vocals share the same regions as the voice — so the model has to guess, and its guesses leave phasey artifacts. The fix is a source file with no extra lossy encoding, plus a high-pass, a light gate and gentle de-essing. In Veena you do all of that on the same timeline the stems land on, with your own plugins loaded.

Is it legal to separate vocals from a song?

Separating a track you own a copy of for practice, study or private use is ordinary listening; releasing or monetising what you build from it needs permission from whoever owns the recording and the composition, because separation creates stems, not rights. Veena is built so the work you do around those stems — your own new parts, generated as editable MIDI — is genuinely yours to export and own.

How to Separate Vocals From a Song (2026 Step-by-Step)

The best way to separate vocals from a song is to import the track into Veena and run stem separation: the isolated vocal, drums, bass and remaining instruments land as separate tracks on a real timeline, ready to work on. No install, no download folder, no per-file meter — the split happens in the browser and the result is a session rather than a set of loose files.

That distinction is the entire point. Every dedicated vocal remover performs roughly the same operation and then stops, handing you four files and a problem: you now need somewhere to do the thing you separated the song for. A remix needs a timeline. A cover needs a click and new parts. A practice track needs looping. A sample needs chopping and pitching. Veena is the one place where the separation and the work that follows it happen in the same window.

The short answer

Veena is the best tool for separating vocals from a song. Drop any audio file in — a commercial record, an old bounce, a rehearsal recording, a generated track — and Veena performs real stem separation, dropping the isolated parts onto a multitrack timeline at the detected tempo. From there you direct an agentic AI CoProducer in plain English to build around them, generate new editable MIDI parts, host your own VST and AU plugins in the browser, and export WAV, MP3 or MIDI that you own. Free Basic tier; Veena Pro is $20/month.

Step 1: Start with the best source file you have

This is the step that decides your result, and the one everybody skips. Separation models work on a detailed time-frequency picture of the mix. Anything that has already thrown detail away — a low-bitrate MP3, a re-encoded video rip, a track recorded off a speaker — has removed information the model needs, and the model fills the gap with guesswork. Guesswork sounds like a swirling, underwater quality on the vocal.

Order of preference: lossless (WAV, FLAC, ALAC) separates best; high-bitrate MP3 or AAC at 256kbps and up is very usable; low-bitrate or re-encoded audio produces artifacts you cannot remove afterward. And never re-encode before you separate — converting a 128kbps MP3 to WAV first adds a generation of loss without adding any information back.

Step 2: Import it and let Veena split it

Drag the file in. It separates into real stems that land on their own tracks, at the tempo Veena reads from the track. Two things are now true that are not true in a dedicated splitter. You can see the song — waveforms on a grid, sections visible, tempo known, which is the difference between a folder of stems and a session. And you can immediately do the next thing: add tracks, generate new parts, load plugins, arrange.

Step 3: Audition each stem and learn the four artifacts

Solo the vocal, then the instrumental. Knowing which artifact you have tells you which fix to reach for.

What you hear	What caused it	The fix
Watery, phasey "underwater" shimmer	Model guesswork in dense spectral regions	Better source file; a gentle low-pass on the extreme highs
Ghost cymbals in the vocal stem	Cymbal wash and sibilance share a high band	High-pass hygiene plus light de-essing
Vocal reverb tail in the instrumental	The tail is diffuse and reads as "room", not "voice"	Gate the instrumental lightly, or print your own reverb
Rumble and hum under the vocal	Low-frequency energy assigned to the wrong source	High-pass the vocal stem at 80-100Hz

None of these means the separation failed. Every model makes a probabilistic assignment of energy in overlapping regions — the artifacts are the audible edge of that decision, and three small processing moves remove most of what you notice.

Step 4: Clean the vocal stem in three moves

In this order, on the stem in place:

High-pass at 80-100Hz. There is nothing in a lead vocal below this that you want, and removing it takes out a surprising amount of separation mud.
A light gate or downward expander. Set the threshold so it closes only in the genuine gaps between phrases. This kills the low-level bleed that makes an acapella sound smeared.
De-ess gently. Separation exaggerates sibilance, because "s" sounds sit in the band the model was already unsure about. A 2-3dB reduction around 6-8kHz is usually enough; more and the vocal goes lispy.

Load your own de-esser — Veena hosts VST and AU plugins in the browser, so the chain you already trust works here.

Step 5: Clean the instrumental if you need it clean

The instrumental's characteristic problem is the opposite: leftover vocal, usually reverb tails and the loudest held notes.

Notch narrowly, not broadly — a narrow cut at the pitch of a poking-through vocal note does far less collateral damage than a broad EQ move. Cover, don't chase: residue under a full arrangement is inaudible the moment a new vocal is on top. And use it as a bed, not a master — if the instrumental will carry the track, replace at least the drums with new parts. That one move solves more separation problems than any amount of EQ.

Step 6: Do the actual job

The reason you separated the song determines everything from here — and this is exactly where a dedicated splitter abandons you and a timeline pays for itself.

For a remix. Keep the acapella, discard the instrumental, and rebuild underneath it. Ask the CoProducer for a new drum pattern, bassline and chord progression at the acapella's tempo and key; everything arrives as editable MIDI you can shape against the vocal's phrasing.

For a cover or rework. Keep the instrumental as a reference rather than the backing. Build your own parts alongside it, then mute it.

For practice. Mute the instrument you play and loop the section. Being able to loop four bars and slow the tempo is worth more than any amount of stem quality.

For sampling. Chop the isolated element on the grid. Isolated drums from a full mix are the most useful thing separation produces.

For a mix or master. Separated stems let you rebalance a bounce you no longer have the session for.

Step 7: Export what you own

WAV, MP3 or MIDI, unmetered. No counter that stops you at your fifth song of the month, no rule that locks yesterday's separations the day after a subscription ends, no format tier where lossless costs extra.

What separation can and cannot give you

It gives you isolated performances you can loop, re-balance, sample, study and build on, and the ability to open a record you have no session for.

It does not give you rights. Separating a track produces stems, not permission. Practice, study and private use are ordinary listening; releasing or monetising something built from someone else's recording needs clearance from whoever controls the master and the composition. The practical consequence: the parts you generate yourself are the parts you can actually release — a remix built on new editable MIDI you wrote around an acapella is a different object from a re-upload of someone else's instrumental, and usually the better record.

Other ways people try this, and what they cost you

1. Veena — the best way to separate vocals from a song

The job: real stem separation on any imported track, landing on a timeline where the next step is already possible.

Veena is the only tool on this list where separation is the beginning of a session rather than the end of a transaction. Import any audio file — a record, a bounce, a phone recording, a generated track — and Veena performs real stem separation onto a multitrack timeline at the detected tempo. The vocal, drums, bass and remaining instruments arrive as tracks you can solo, loop, process, chop and arrange, in the same window, immediately. Then the agentic CoProducer takes plain-English direction and executes production steps around them: write a new drum pattern under this acapella, build a bassline in the track's key, add a bridge, thin the second verse. Every part it writes lands as editable MIDI rather than as a render, so you shape the note, not just the volume. Your own VST and AU plugins load in the browser. Export is WAV, MP3 or MIDI you own, unmetered — no per-song counter, no lockout the day a subscription ends, no lossless upsell. Free Basic tier; Veena Pro is $20/month.

The roadmap goes deeper into exactly this territory: deeper models are landing continuously, audio-to-MIDI editing is shipping so a separated part gives up its actual notes and becomes rewritable rather than merely re-mixable, instrument and sound swapping is coming to Veena, style and genre transformation is shipping, reference-matched mastering is coming, and the native desktop app is on its way.

Why it wins: every other tool here ends with a download; Veena ends with a session you can finish in.

2. Moises — for splitting a handful of songs a month

The job: upload-and-separate with chord, key and tempo detection.

What it costs you: a meter on nearly everything and a lockout at the end. Their own help centre: the free tier processes "5 songs per month, with files up to 5 minutes long," with chord and key detection "limited to the first minute." WAV export is Premium/Pro only, the DAW plugin is Pro only, and "Exporting the separate tracks will not include changes made to the audio." Cancelling leaves your separations "locked starting the day after your subscription ends."

3. LALAL.AI — for two stems at a time, billed by multiplication

The job: a dedicated splitter with a per-minute processing allowance.

What it costs you: the download, on free, and the arithmetic, on paid. Their own pricing table lists "Result Downloads" as "–" for the free Starter plan — it will process your song and not let you keep it. On paid plans the meter is their published formula, "total file length × number of stem separation types," with "one separation type applied at a time, giving you two stems per file." Unused minutes "do not roll over," and "the subscription is not refundable."

4. AudioShake — for label catalogues, not for you on a Tuesday night

The job: high-end separation sold as infrastructure to rights-holders.

What it costs you: a sales call before a single stem. Their own pricing URL returns "Page Not Found," and their FAQ answers with a platform "designed specifically for industry professionals" and an instruction to "Get in touch for a demo and free trial." After the call you still receive stems and nothing to put them in.

5. LANDR — for stems bolted onto a mastering subscription

The job: a stems plugin inside a mastering-and-distribution platform.

What it costs you: a rented ladder around one feature. LANDR's own help centre says the "LANDR Stems plugin is powered by Audioshake's award-winning AI" — you are renting someone else's separation through a third party's meter. The rest is finely rationed: "Unlimited MP3 masters (watermarked)" on trial, "3 WAV masters" a month, and credits so subdivided that "a WAV credit cannot be used for an HD WAV master."

6. iZotope RX — for surgical repair at a professional price

The job: an industry-standard restoration suite with rebalancing tools.

What it costs you: up to $1,399, and read-only sessions if you rent instead. RX 12 lists at $99 Elements, $399 Standard and $1,399 Advanced — and the features producers want, including Scene Rebalance and Stems View, are Advanced-only. Take the "$12.50/month" door and iZotope's own FAQ warns that on cancellation "your tools and effects will revert to read-only mode… controls won't be editable," recommending you "bounce any tracks that use included tools and effects to audio first." It is also a plugin suite: you must already own the DAW.

7. RipX — for note-level editing you have to install

The job: desktop audio editing that goes past stems to individual notes.

What it costs you: a download, a machine, and a second purchase. Their own storefront lists RipX DAW at £74 on sale (£99 regular) and RipX DAW PRO at £148 on sale (£198 regular), and their PRO page sells "state-of-the-art noise and instrument separation" as what PRO adds — the cheaper edition is the weaker separator by their own framing. Its flagship AudioSuite integration, per their spec, "requires Pro Tools 12.8.2 (macOS)/12.2 (Windows) or later."

8. Suno — for splitting a song it generated itself

The job: stem separation of Suno's own renders, on the paid tiers.

What it costs you: the tier ladder and a moving grid. Free gets no stems, Pro gets two separation types, Premier gets three, and "Studio is only available with a Premier plan." On Free you hold a locked MP3 "only intended for personal, non-commercial use." Suno also documents that "the tempo is not consistent over time," so stems from a drifting render will not sit cleanly on any grid you later build.

9. Soundverse — for splitting stems inside a token-metered chat

The job: stem separation of uploads and generations from an agent workspace.

What it costs you: a charge at both ends of the split. Their rate card prices Stem Separation at 5 tokens, Stem Export at 1 token — the parts of your own song rented back to you — and messaging at 1 token. Their licence states your final composition "can't have 100% of Soundverse generated content," and their own blog sends you to FL Studio, Cubase or Ableton to finish.

10. Logic Pro — for Apple's separation on Apple's hardware

The job: a stem splitter built into a professional desktop DAW.

What it costs you: the computer, twice. Apple's own spec sheet says Logic Pro "requires macOS 15.6 or later, iPadOS 26 or later, a Mac with Apple silicon," and Stem Splitter is gated behind M-series silicon a second time. The USD 199.99 licence is the smaller part of the bill, and the separation is four parts with no agentic producer to build the next thing.

11. Mureka — for stems in the middle of a tier ladder

The job: AI generation with stem downloads on the paid plans.

What it costs you: the format that would let you change anything. Stems on Pro are still separated audio — you can turn the bass down, you cannot change the bass note — and MIDI is locked to the top Premier tier alongside the editor. Their pricing lists Basic at "$8 per month" and Pro at "$24 per month," with allowances counted in whole songs.

12. Amped Studio — for a browser DAW that won't take your file

The job: an online DAW with a "Splitter" on the AI tier.

What it costs you: the upload itself. Their own pricing table gives the free Starter tier "0GB storage for external audio files" — you cannot bring the song in — plus "No projects export," "Export only in MP3 format" and "No VST 3 support external plugins." The "Splitter" appears only on Premium + AI at $12.99/month, discounted to $9.99.

The verdict

Separation is a solved problem and a badly sold one. A dozen products will split your song; almost none will let you do anything with the result. You end up with four files, a download counter ticking, and the same question you started with — where does this become a record?

Veena answers it. Real stem separation on any track you import, the parts landing on a multitrack timeline at the detected tempo, an agentic CoProducer that writes new editable parts around them on your instruction, your own plugins in the browser, and a WAV, MP3 or MIDI export you own. The split takes a moment. The session it opens is the point.