AI Vocal Remover — Isolate Vocals with AI
An AI vocal remover does not subtract a channel — it listens to the whole mix and predicts which part of the sound belongs to the singer. That is why the result sounds like a performance instead of a hollow karaoke file.
Upload an MP3, WAV, M4A or FLAC below and get vocals, instrumental, drums, bass and other stems back in minutes. Runs are paid: the one-time $2.90 Trial Pack includes 15 minutes and the full 4-stem output.
By StemSplit Audio TeamLast updated: Editorial standards
Trial Pack: 15 minutes for $2.90, one-time payment. Runs start after checkout.
Paid runs only · $2.90 Trial Pack includes 15 minutes · files deleted automatically after processing
Four stems, one upload
Every job returns the full set of separations, not just a karaoke-style instrumental. That means one upload gives you an acapella, a backing track, a drum loop and a bass line you can drop straight into your DAW.
What comes back in your download
Lead and backing vocals, harmonies and ad-libs, isolated so you can re-sing them, sample them or build an acapella.
Kick, snare, hats and percussion as one clean stem for remixes, drum practice and loop-based production.
Bass guitar and sub-bass separated from the mix, ready for sampling, transcription or replacing the low end.
Guitars, keys, strings, synths and background textures — every element that is not vocals, drums or bass.
The instrumental is simply your original track with the vocal stem removed — it is created in the same pass, so you get vocals, instrumental and the individual stems from a single upload.
Why AI separation beats the old center-channel trick
The classic vocal remover works by phase-cancelling the center of a stereo mix. It inverts one channel and sums it with the other, so anything identical in both channels disappears — usually the lead vocal, but also the kick, snare, bass and anything else the mix engineer placed in the middle. What is left sounds thin and hollow, and it falls apart completely on vocals that are wide, doubled or panned.
An AI vocal remover approaches the problem from the other direction. The model has learned what vocals, drums, bass and other instruments sound like across thousands of songs, so it can estimate each source's contribution at every frequency and rebuild it as a separate signal. The backing track keeps its punch, and the vocal keeps its detail.
- Center-cancel tools remove centered drums and bass along with the vocal.
- Phase tricks fail on wide, doubled or reverberant vocals.
- AI separation keeps the full band and returns individual stems.
- The same upload also produces an acapella for covers and remixes.
Inside the model: what htdemucs actually does
StemSplit runs a Demucs-family model (htdemucs, originally developed by Meta AI and released under an MIT license). Demucs is a hybrid transformer: it looks at the raw waveform and at a time-frequency representation of the audio at the same time. The waveform view preserves transients — the attack of a snare, the plosive in a vocal — while the spectrogram view makes it easier to separate sustained tones and overlapping harmonies.
The model outputs four source estimates: vocals, drums, bass and other. Those estimates are mixed back together into the original song, and during training the model is optimised to make that reconstruction as close to the real mix as possible. The practical effect is that the four stems sum back into something that sounds like the record, while each file is usable on its own.
The model does not learn from your upload. Your track is processed once, the stems are rendered, and the audio is deleted from working storage afterwards. Nothing you upload becomes training data.
AI vocal remover vs. EQ and phase cancellation
If you only need a rough karaoke backing track and you already use a DAW, the phase trick can be good enough — especially on simple stereo recordings where the vocal is dead center. It takes seconds and adds no cost. But it is a blunt instrument: it cannot tell a vocal from a bass guitar if both are centered, and it removes the low end of the mix along with the singer.
For acapellas, sampling, remixes or anything you plan to publish, AI separation is the only option that produces a full-spectrum result. The trade-off is processing time and cost: the $2.90 Trial Pack buys 15 minutes of full 4-stem separation, which is roughly five three-minute songs. Use phase cancellation for quick scratch work and AI separation for anything a listener will hear.
- Phase cancellation: instant and zero-cost, but loses centered instruments and struggles with stereo vocals.
- EQ carving: removes a frequency band, not an instrument — vocals and other midrange content both suffer.
- AI separation: slower, but keeps the full arrangement and returns individual stems.
Quality depends on the source file — how to help the model
No separator can recover detail that is not present in the file. A 128kbps stream rip, a clipped phone recording or a heavily limited master gives the model less to work with, and artifacts become more likely. The single biggest improvement you can make is uploading the highest-quality copy of the track you have.
Trim silence and count-ins before uploading where possible. Shorter clips process faster and use less of your paid minutes. If you are working with a live recording, expect bleed between stems — a vocal mic that captured the drums will put some drums in the vocal stem, and no model currently solves that perfectly.
- Prefer WAV or 320kbps MP3 over low-bitrate streams.
- Avoid files that clip or peak above 0dBFS.
- Trim long intros and outros to save paid minutes.
- Live and mono recordings are harder than studio mixes — expect some bleed.
Where the AI still struggles
AI vocal removal is excellent on modern studio productions and good on most older stereo recordings. It is weaker on mono masters from the 1960s, dense distorted mixes where every instrument occupies the same frequencies, orchestral recordings with many overlapping sources, and spoken-word tracks where speech and background noise are entangled.
When the result is not clean enough, try the high-quality model on Pro, process a shorter section, or use the stems for editing rather than as a final master. For remix and karaoke work, small artifacts are usually masked by the new arrangement; for critical listening, always compare the result against the original.
Trial Pack, Starter or Pro: what to buy
There is no free-of-charge allowance — every run is paid, which keeps the queue short and the model quality consistent. The entry point is the one-time $2.90 Trial Pack: 15 minutes, 4 stems, MP3 320 and WAV, no subscription. Fifteen minutes covers roughly five three-minute songs and never expires.
If you process more than that, Starter at $5/month includes 90 minutes with batch upload; Pro at $9/month raises it to 200 minutes and adds the 6-stem model, high-quality mode and a priority queue. The annual plan is $39 — five months of Starter for twelve months of access.
Sample outputs are illustrative; real A/B samples will be added after launch. No customer counts, reviews or benchmark numbers are published on this page because we have not measured them yet. Every upload is processed only for your job and removed from working storage automatically within 24–48 hours — read the privacy policy.
System built-in method: Windows, macOS, phone and DAW
Your OS and DAW can already perform one kind of vocal removal: center-channel cancellation. Knowing what that built-in method actually does is the fairest way to judge what an AI model adds — and where it still fails.
Windows 11 / 10
Windows has no built-in vocal remover. Voice Recorder and Clipchamp can record, trim or change volume, but they cannot separate or cancel a vocal.
- Install Audacity from the official site (audacityteam.org). It is free and open source, but it is not bundled with Windows.
- Open your song as a stereo track (File → Open). Mono files have no center channel to cancel.
- Select the whole track (Ctrl+A), then open Effect → Vocal Reduction and Isolation.
- Choose Action: Remove Vocals, keep the default settings and apply.
- Check the result on headphones: centered bass and kick are usually removed together with the vocal.
- Export with File → Export Audio and choose WAV or MP3.
Good for a rehearsal backing track. It cannot keep the low end and it cannot produce an acapella or stems.
macOS
macOS ships GarageBand, but GarageBand has no phase-invert plug-in — so the bundled app alone cannot cancel a centered vocal. QuickTime Player can trim and export audio, not separate it.
- Open the song in GarageBand (bundled) if you only need to trim, split or adjust level — that works.
- GarageBand's EQ can notch 1–4kHz to reduce vocal presence, but it also thins guitars, cymbals and snare.
- For actual center cancellation, use Audacity (free, cross-platform) and follow the Windows steps above.
- Logic Pro (paid, not bundled) includes the Gain plug-in with Phase Invert: split stereo, invert one side, then sum to mono.
- Listen in mono afterwards — phase tricks can sound different when a phone speaker folds the mix down.
Honest answer: macOS has no bundled vocal remover. GarageBand EQ is a partial workaround, not a separator.
iPhone & iPad
iOS and iPadOS have no built-in vocal remover for your own files. GarageBand for iOS can record and arrange, but it has no phase-invert control. Apple Music's Sing feature can lower vocals on supported catalog songs, but it streams only and cannot export an instrumental.
- Apple Music Sing: play a supported song, open the lyrics view and use the vocal slider to lower the vocal (Apple Music subscription required).
- There is no export: Sing is a playback feature, not a file tool, so it cannot give you a WAV or MP3 backing track.
- GarageBand for iOS can trim and arrange audio you import, but it cannot cancel or remove a vocal.
- iMovie can lower a clip's volume, not isolate a source.
- For real separation, open this site in Safari on the phone, or move the file to a Mac/PC and use a desktop method.
Phones are for listening and file handling. No iOS feature produces stems or a downloadable instrumental.
Android
Android has no system vocal remover. Files by Google, Google Photos and Google Drive can move or trim media, but none of them separate audio sources.
- Google Files and Drive can move the audio between apps, but cannot edit or separate it.
- Third-party Android 'karaoke' or 'vocal remover' apps usually run the same center-channel math and damage the mix.
- If an app asks you to upload the track to its server, read its privacy policy first — many keep files far longer than they say.
- For AI separation, open this site in Chrome on Android; the upload and download flow works in the browser.
- For manual phase cancellation, move the file to a desktop with Audacity.
Android's built-in apps do not separate stems. Treat app-store vocal removers as phase filters unless they clearly run an AI model.
DAW (Audacity, GarageBand, Logic Pro, REAPER)
Every DAW can perform center-channel cancellation manually if it supports channel splitting and phase inversion. This is the real built-in technique — no plug-in purchase required.
- Import the song as a stereo track and split it into two mono channels.
- Invert the phase (polarity) of one channel only.
- Sum both channels down to mono. Anything centered — usually the lead vocal — cancels out.
- In REAPER, use the item phase (polarity) button before summing; in Logic Pro, use the Gain plug-in's Phase Invert.
- Audacity's Effect → Vocal Reduction and Isolation automates the same math.
- Check the result on headphones: if the bass and kick vanished, the method worked as designed — that is its limit.
Instant and offline, but the output is a mono backing track, not stems. There is no acapella and no preserved low end.
Built-in method vs. AI separation
Built-in cancellation throws away centered drums and bass along with the vocal, and it cannot return an acapella or separate stems. The AI model estimates each source across the frequency range instead; the one-time $2.90 Trial Pack includes 15 minutes to compare on your own track.
Documentation we check for these steps: Audacity manual · Apple GarageBand · Apple Logic Pro · REAPER · Demucs (Meta AI) on GitHub · Replicate: cjwbw/demucs.
Frequently asked questions
Can I try before paying?
No. There is no free-of-charge processing tier and no complimentary preview — every separation costs GPU time, so every run is paid. The cheapest entry point is the one-time $2.90 Trial Pack: 15 minutes, 4 stems, WAV + MP3, no subscription.
Is AI vocal removal legal?
Separating a song you own or have licensed is generally fine, and we grant you a commercial-use license for the processing service. We cannot grant rights to the underlying recording. Uploading someone else's copyrighted song and publishing the stems may infringe their rights, so only process audio you are allowed to use.
Can it isolate backing vocals separately from the lead?
Not on the standard 4-stem model. Everything sung — lead, harmonies and backing vocals — lands in the vocals stem, which is exactly what you want for an acapella. The Pro plan's 6-stem model adds more instrument separation, but lead-versus-backing vocal splitting is not offered yet.
Does the AI vocal remover work on any genre?
It works best on studio recordings with a clear arrangement — pop, rock, hip-hop, EDM, country, R&B and karaoke tracks. Dense metal, orchestral music, live bootlegs and mono masters are harder and may show more bleed between stems.
What is the difference between an AI vocal remover and stem separation?
They are the same engine described from different angles. 'Vocal remover' focuses on the vocal-versus-instrumental result; 'stem separation' focuses on the four individual parts. Every job here returns both: an acapella, an instrumental, drums, bass and other.
Can I get six stems instead of four?
The Pro plan includes the 6-stem model in addition to the standard 4-stem output. It separates the 'other' bucket further for more detailed production work. The Trial Pack and Starter return the standard 4 stems.
Hear what the AI does with your own track
Buy the $2.90 Trial Pack, upload up to about five songs and compare the acapella and instrumental with the original. 15 minutes, one-time payment, no subscription.
Read how the stem separation works, how your audio is handled, or contact the team.