creative

Best AI Vocal Removers

AI vocal removers and stem splitters ranked on independent listening tests, published separation scores, user ratings, free tiers and price, plus the DAW and DJ software with built-in stems and what the law says about splitting songs you do not own.

Updated October 2026

Quick answer: The best AI vocal remover for most people is LALAL.AI, because it is the only browser tool near the top of both independent multi-tool stem separation tests, finishing third of 11 behind two desktop applications in MusicRadar’s test and standing out on drums in MusicTech’s test, and it holds 4.4 from about 4,330 Trustpilot reviews. For musicians who want to practise along to songs on a phone, Moises is the better choice, because it adds tempo, pitch and chord tools around the separation and has 4.7 from about 28,000 App Store ratings. For the cleanest vocals at no cost, MVSEP gives 50 free separations a day and was MusicTech’s pick for vocals, and Ultimate Vocal Remover is the best free desktop app. If you already own a Mac with Apple silicon, Logic Pro’s Stem Splitter won MusicRadar’s test outright. The caveat that applies to every tool: no separator is lossless on every song, reverb and backing vocals leak between stems on all of them, and the only published separation scores for commercial services predate the models they sell today.

This page ranks tools that split an existing recording into vocals, instrumental and individual stems such as drums, bass, guitar and piano. It is a different job from the pages next to it:

• Best AI music generators ranks tools that create new songs, such as Suno, from a prompt.
• Best AI voice cloning tools covers tools that make a synthetic voice speak or sing.
• This page ranks separation of audio you already have: karaoke tracks, acapellas for remixes, practice stems and samples.

Which one should you pick?

What the evidence says about AI stem separation

Separation quality is measured in signal-to-distortion ratio (SDR), in decibels. A higher SDR means the separated stem is closer to the original isolated recording. The standard academic test set is MUSDB18-HQ, 150 songs with their original multitrack stems.

The best academic models reached about 12 dB on vocals by 2023, and progress since has been incremental. ByteDance’s BS-RoFormer scored 12.72 dB on vocals and an average of 11.99 dB across four stems on MUSDB18-HQ when trained with extra data (Lu et al., arXiv, September 2023). Meta’s HT Demucs v4, the model inside many free tools, averaged 9.20 dB with extra training data (Rouard et al., arXiv, November 2022). The newer SCNet-large reached 11.10 dB on vocals with extra data (arXiv, January 2024).

The largest public leaderboard is run by a company that sells separation. MVSEP’s Multisong leaderboard scores any model submitted against its own song set and recalculates every three hours. At the top, MVSEP’s own BS-RoFormer model from July 2026 scored 12.33 dB on vocals and 18.64 dB on instrumentals; the best entry not made by MVSEP, ByteDance’s sami-bytedance v1.1, scored 11.82 dB and 18.13 dB. MVSEP sells credits for the same models, and one of its developers, Roman Solovyev, won first prize in the 2023 Sound Demixing Challenge, so its leaderboard is useful but not independent of the company.

The only published scores for commercial services are from February 2023. On the MVSEP leaderboard, community member Bas Curtiz submitted LALAL.AI’s Phoenix model at 7.47 dB on vocals, Splitter.ai at 10.57 dB and Audionamix XTRAX Stems at 5.58 dB, all in February 2023. LALAL.AI has since replaced Phoenix with two newer models, Perseus and Andromeda, so these figures do not describe what the services sell today. No current independent SDR measurement of LALAL.AI, Moises, AudioShake, Logic Pro or any other commercial tool has been published.

SDR does not reliably predict which separation sounds best. In the 2023 Sound Demixing Challenge, ByteDance’s SAMI team had the highest score on the standard leaderboard (9.97 dB mean, 11.36 dB on vocals), but seven expert listeners making 583 comparisons ranked the top three systems in a statistical tie, and the organisers reported no clear correlation between human judgement and SDR (Fabbro et al., arXiv). That is why this page weighs listening tests as heavily as scores.

The one large human rating of commercial tools found them close together. In the paper introducing its own SAM Audio model, Meta had listeners rate separations from 1 to 5 (Meta, arXiv, December 2025). On professionally produced instrument material, AudioShake scored 4.28, open-source Demucs 4.26, Moises 4.22, LALAL.AI 4.18, Fadr 3.92 and Spleeter 3.90. On in-the-wild material the gaps widened: AudioShake 3.43, Moises 3.12, LALAL.AI 3.03 and Fadr 2.45. Meta’s own model won every category in Meta’s own test, so its figures are reported as Meta’s claim, and the comparison among the other tools is the useful part.

Independent evidence at a glance

How we judged these tools

A tool is ranked here only when there is independent evidence about it from more than one source: a hands-on listening test by a publisher that does not sell a separator, a published benchmark or human rating, and user ratings with real volume behind them. Vendor benchmarks, vendor blog rankings and paid press releases are reported as the company’s view and never counted. The full method is on how we rank.

Two independent multi-tool tests carry most of the weight. MusicRadar’s Stuart Adams ran three mixed songs through 11 tools at their highest quality settings, scored the top five out of 20 and used phase-cancellation null tests to check whether the stems summed back to the original. MusicTech’s Alex Holmes ran one track through nine tools and judged each stem type for leakage, timbre, reverb tails and artefacts, without naming an overall winner. Neither test was blind, and MusicTech’s page carries an affiliate disclosure.

Vendor-run rankings are reported but not counted. AudioShake’s comparison page claims 13.5 dB on vocals against 8.9 dB for Demucs v4 without naming the test set. Music.AI’s benchmark post (17 September 2024) claims an average SDR 15.8% above its nearest competitor on MUSDB18-HQ and an in-house set, with competitors unnamed. LALAL.AI says its Andromeda model improves SDR by “up to 10%” over Perseus, without a named test set (LALAL.AI).

A community test is reported but not counted. Music By Mattie (3 April 2026) scored 12 tools on 13 songs and placed Ultimate Vocal Remover first (8.05), Moises second (7.85) and Ableton Live third (7.20), with LALAL.AI fifth (6.20). The test was not stated to be blind and its author sells presets and plug-ins, so we report it as context.

Best AI vocal removers ranked

The best AI vocal removers, reviewed

1. LALAL.AI — best AI vocal remover overall

Made by: OmniSale GmbH Current model: Andromeda (web); Lyra (offline in the VST plug-in) Price: Free (10 minutes in the slow queue); Lite $7.50/month or $90/year (90 fast minutes a month); Pro $15/month or $180/year (250 fast minutes, plug-in and API); one-off packs from $50 for 750 minutes Stems: Vocals, instrumental, drums, bass, guitar, synth, strings and wind; separate lead and backing vocal splitter

LALAL.AI is the best vocal remover for most people because it is the strongest tool in both 2026 independent tests that runs in a browser with no install. In MusicRadar’s test it placed third of 11 on 15/20, behind only two paid desktop applications, and Stuart Adams wrote that “the quality of results delivered by LALAL.AI is undeniable”, with vocals, drums and bass stronger than its extended instrument stems. In MusicTech’s test, Alex Holmes placed its vocals in the close leading group, named it “the standout in terms of transient punch and consistency” on drums and placed it a close second on bass. LALAL.AI holds 4.4 from about 4,330 reviews on Trustpilot, where the company invites reviews, and 4.4 from 647 ratings on the App Store.

What it does. LALAL.AI’s web tool removes vocals in one step and splits up to eight stem types in its Stem Splitter, according to its pricing page. LALAL.AI says its Andromeda model is up to 40% faster than Perseus with “up to 10%” better SDR (LALAL.AI); no independent test has measured that. The Pro plan adds a VST plug-in that runs six stems offline on a local model called Lyra, according to Mixdown (28 April 2026).

Watch out for: Minutes are charged per stem, so splitting a five-minute song into four stems uses 20 minutes, according to LALAL.AI’s pricing FAQ. The free tier is 10 minutes in the slow queue and a 200 MB file limit, which is enough to test quality but not to use it regularly. MusicRadar criticised the interface and its one-stem-at-a-time workflow. LALAL.AI says it does “not use user files for artificial intelligence training” (privacy policy).

Why it wins over Moises. The two are close, and Meta’s human ratings put them within 0.04 of each other on professional material. LALAL.AI wins for someone who searches for a vocal remover because both 2026 listening tests rated its vocals better: MusicRadar placed it third overall, while Moises missed the top five for “warbling, clicks and spill” in its vocal stems. Moises has far more user ratings and the better free allowance, which is why it is the pick for practising musicians.

2. Moises — best for practising musicians and mobile

Made by: Music AI Inc. (Moises is the consumer app; Music.AI is the same company’s business API) Price: Free (5 uploads a month); Premium $5.99/month or $39.99/year; Pro $29.99/month (App Store prices) Stems: Vocals, instrumental, drums, bass, guitar, piano and more; Moises claims 27 stem types

Moises is the best choice for musicians who want to learn, practise or sing along, because it wraps separation in tools for slowing a song down, changing its key, detecting chords and adding a click track, according to its App Store listing, and it has the largest body of user ratings in the category: 4.7 from about 28,000 ratings on the US App Store and between 4.5 and 4.6 from about 430,000 reviews on Google Play. In Meta’s December 2025 human ratings Moises scored 4.22 on professional instrument material, and in the music category it beat AudioShake 3.90 to 3.37 (Meta, arXiv).

What independent testers found. MusicRadar found its drum and bass extraction good but said it was “let down by its vocal extraction”, with warbling, clicks and spill from other stems. MusicTech found its vocal separation “very slightly more aggressive”, with less even reverb tails, a result it shared with Ableton Live, whose stem separation MusicTech reports is also powered by Music AI.

Company. Music AI raised a $40 million Series A on 22 January 2025, when Moises had 50 million registered users, according to Music Business Worldwide.

Watch out for: Deleted songs still count towards the five free uploads a month, according to Moises help. The highest-quality “Hi-Fi” models are on the Pro tier only. Moises refuses links from streaming services, on the grounds that extracting stems from them would be “an unauthorized derivative work”, and tells users to process only music they hold rights to (Moises help).

3. Apple Logic Pro Stem Splitter — best for Mac producers

Made by: Apple Current version: Logic Pro 12 (released 30 January 2026); six-stem Stem Splitter since Logic Pro 11.2 (28 May 2025) Price: $199.99 one-off, no subscription Stems: Vocals, drums, bass, guitar, piano, other

Logic Pro’s Stem Splitter is the best choice for anyone who already produces on a Mac, because it won MusicRadar’s test outright on 16/20, with Stuart Adams writing that “recognition accuracy is impeccable and stem extraction is very clean”. MusicTech placed its vocals in the leading group and called it “a standout” for piano, guitar and other instruments. In Sound On Sound’s review of Logic Pro 11.2, John Walden was “generally impressed” and found the six-stem version cleaner than the original, while noting that vocals suffer where reverb or harmonies are present.

What it does. Stem Splitter runs on the Mac itself, so audio never leaves your computer, and places each stem on its own track ready to edit. Apple added guitar and piano stems in Logic Pro 11.2 on 28 May 2025, according to Apple. Logic Pro 12 went on sale on 30 January 2026 at $199.99, according to Synthtopia.

Watch out for: Stem Splitter needs a Mac with Apple silicon; Sound On Sound states at least an M1. There is no Windows version. For vocal removal alone, LALAL.AI and MVSEP run in a browser with no purchase.

4. MVSEP — best for the cleanest free vocals online

Made by: MVSep.com Current models: BS-RoFormer, Mel-RoFormer, SCNet and others, including MVSEP’s own trained versions Price: Free (50 separations a day, according to MusicTech); credit packs from $3 for 60 credits to $1,000 for 110,000 credits, no subscription (MVSEP)

MVSEP is the best choice for clean vocals at no cost, because MusicTech’s Alex Holmes named its BS-RoFormer vocal model the best for smoothness and consistency in a nine-tool test, and named it the cleanest and best at identifying parts in the other-instruments stem (MusicTech). Its model family also leads the published scores: BS-RoFormer is the architecture with the highest MUSDB18-HQ vocal score in the academic literature (12.72 dB), and MVSEP developer Roman Solovyev won first prize in the 2023 Sound Demixing Challenge.

What it does. MVSEP lets you pick from dozens of separation models, including vocal, instrumental, drum, bass, guitar, piano and reverb-removal models, and run several in sequence. Its own leaderboard scores its July 2026 BS-RoFormer model at 12.33 dB on vocals and 18.64 dB on instrumentals (MVSEP).

Watch out for: MVSEP runs the leaderboard that its own models top, on a song set it chose, so its scores are the company’s figures rather than independent ones. The interface is built for enthusiasts: model names such as “bs_roformer” assume you know what you are choosing. MusicRadar did not test it, so it has one independent hands-on test where LALAL.AI has two.

5. Ultimate Vocal Remover — best free desktop app

Made by: Anjok07 and contributors (open source, MIT licence) Current version: 5.6 (stable, 2023), with a separate beta build that adds Roformer and SCNet models Price: Free

Ultimate Vocal Remover is the best free vocal remover you can run on your own computer, because MusicRadar placed it fifth of 11 on 13/20 and described its MDX-Net vocal separation as “exceptional and entirely lossless” (MusicRadar). It is a graphical front end for the same open model families scored in the academic literature: MDX-Net, Demucs and, in the beta build, BS-RoFormer and SCNet.

What it does. Ultimate Vocal Remover runs on Windows, Mac and Linux, processes files locally with no upload and no limits, and lets you combine models into ensembles. With Demucs models it splits up to six stems.

Watch out for: There has been no new stable release since version 5.6, which file dates on its releases page place in 2023, and the newest Roformer models are only in a beta build, according to its GitHub releases page. MusicRadar found its Demucs mode weaker on bass than its MDX-Net vocals. Separation is slow without a capable graphics card. Meta archived the original Demucs repository on 1 January 2025, so Demucs models are no longer developed by Meta (GitHub).

6. Steinberg SpectraLayers Pro — best for audio post-production

Made by: Steinberg Current version: SpectraLayers 13 (released 6 July 2026) Price: Pro $359.99; Elements $89.99 (Sonic State) Stems: Unmix Song separates vocals, drums, bass, sax and brass, guitar and piano; Unmix Drums splits six kit pieces

SpectraLayers Pro is the best choice when you need to edit what the AI separated, because it shows the audio as a spectrogram you can paint on and move between layers, and its separation placed second in MusicRadar’s test on 15/20. Stuart Adams found its processing lossless but its recognition “isn’t infallible”: it put a piano in the guitar stem on one song and cleanly extracted a bass synth on another. Sound On Sound rated SpectraLayers 11’s slowest Extreme mode above Logic’s splitter, at the cost of much longer processing.

What it does. SpectraLayers 13 added tools aimed at film and podcast work, including Unmix Sound Effects, Unmix Two Voices and Voice DeCrosstalk, according to Sonic State.

Watch out for: MusicTech found its drum stems “a little messy”. Check which separation tools the $89.99 Elements edition includes before buying it.

Pricing comparison

Minutes and credits are not comparable across tools. LALAL.AI charges per stem, so a four-stem split costs four times the song’s length; Moises counts uploads; MVSEP charges credits that vary by model.

Best AI vocal remover for each job

Best free AI vocal remover

MVSEP is the best free online vocal remover because it allows 50 separations a day and MusicTech rated its vocals the best in its test. If you would rather not upload anything, Ultimate Vocal Remover is free with no limits on your own computer. Moises’s free plan allows five uploads a month, and LALAL.AI’s allows 10 minutes in total.

Best AI vocal remover for karaoke

LALAL.AI is the best choice for making karaoke instrumentals, because a two-stem vocal removal costs one minute per minute of song and MusicRadar ranked its vocal separation among the best. For singing along on a phone with lyrics and key changes, Moises is the better app.

Best AI stem splitter for producers and remixers

Apple Logic Pro is the best stem splitter for producers on a Mac because it won MusicRadar’s test and puts six stems straight onto tracks in the DAW. On Windows, LALAL.AI’s Pro plan adds a VST plug-in that separates six stems offline inside any DAW.

Best AI stem splitter for drums and bass

LALAL.AI is the best choice for drum and bass stems, because MusicTech named it the standout on drums and a close second on bass, and MusicRadar found its drums and bass stronger than its other instrument stems.

Best AI vocal remover app for iPhone and Android

Moises is the best vocal remover app for phones because it has 4.7 from about 28,000 App Store ratings and about 430,000 Google Play reviews, the largest body of user ratings in the category.

Best AI vocal remover for privacy

Ultimate Vocal Remover and Logic Pro are the best choices for unreleased or confidential audio, because both separate on your own computer and never upload the file.

Stems in DAWs and DJ software

Many people already own a separator inside software they use for something else.

DAWs. Ableton Live 12.3 added four-stem separation (vocals, drums, bass, others) to Live Suite in November 2025, running locally, and MusicRadar placed it in the “closely-fought middle ground” with FL Studio. FL Studio has had four-stem separation since version 21.2 on 30 October 2023 (Mixmag). MusicRadar found Cubase Pro 15’s stems “heavily processed” and Studio One Pro 7’s prone to spill. Bitwig said on 13 October 2026 that stem separation is coming to all editions of Bitwig Studio “this winter” (Sonic State). iZotope RX 12, released on 29 April 2026 from $99, adds a Stems View built on its Music Rebalance tool; MusicRadar found the previous version’s stems “more processed”.

DJ software. Real-time stems let DJs mute vocals or drums live. In Digital DJ Tips’ listening test (15 January 2025), djay Pro sounded best, followed by VirtualDJ, Traktor, Serato and rekordbox, with Engine DJ described as “warbly” and “bleedy”. Rekordbox added a fourth stem in version 7.2.8 on 3 December 2025, which Digital DJ Tips still judged “lagging behind the best-in-class” (Digital DJ Tips). djay Pro has 4.7 from about 76,000 ratings on the US App Store; its Neural Mix stems do not work on Apple Music or Spotify streams.

Other AI vocal removers worth knowing (not ranked)

These tools are not ranked. Each description below is taken from the company’s own pages, and its claims are the company’s claims.

• LANDR Stems is LANDR’s stem separator, powered by AudioShake. It placed fourth in MusicRadar’s test on 13/20 and was MusicTech’s pick for bass. It runs as a DAW plug-in and is sold as part of LANDR Studio subscriptions rather than on its own (Sound On Sound).
• AudioShake sells separation to businesses through AudioShake Studio, which it describes as “free to start”, and a per-minute API, with ESPN, Warner Bros and Universal Music Group named as customers. It raised a $14 million Series A on 13 November 2025 (Slator). It scored highest of the commercial tools in Meta’s December 2025 human ratings, but it has no consumer plan with a published price.
• Fadr offers unlimited free separation into vocals, melodies, drums and bass, with MP3 output and 10-minute uploads; Fadr Plus costs $10 a month or $100 a year for 18 stem types, WAV output and plug-ins (Fadr). It holds 4.6 from 196 Trustpilot reviews, and Meta’s human ratings placed it below LALAL.AI and Moises.
• Gaudio Studio offers 20 free minutes with MP3 output, then minute packs from $7 for 50 minutes. MusicTech placed its vocals in the close leading group and found its drums “set very loose”.
• vocalremover.org is a free, donation-funded browser tool that returns a karaoke track and an acapella.
• Acon Digital Remix is a $49 plug-in that splits five stems (vocals, piano, bass, drums, other) in real time inside a DAW.
• RipX DAW is a desktop audio editor built around separated notes, from $99 for RipX DAW and $198 for RipX DAW PRO, with version 8 announced on 11 December 2025 (MusicTech).
• Suno can split a song made in Suno into 2 or 12 stems on its Pro and Premier plans. It is a stem export for Suno’s own output, covered on best AI music generators.

Left out: Deezer’s Spleeter, a 2019 open model that newer models score well above on MUSDB18-HQ.

Why every vocal remover still leaves artefacts

All six ranked tools leave some sound behind in the wrong stem, and the independent tests describe the same failure points.

• Reverb and delay. Reverb tails belong to both the vocal and the room, so separators split them unevenly. MusicTech found uneven reverb tails on Moises and Ableton; Sound On Sound found Logic’s vocals suffered with reverb present.
• Backing vocals and harmonies. Stacked voices blur together. LALAL.AI sells a separate lead and backing vocal splitter for this reason.
• Instruments in the same range as the voice. Guitar, piano and synth parts that overlap the vocal are the most often misclassified. MusicRadar found SpectraLayers placed a piano in the guitar stem.
• Low-quality sources. A 128 kbps MP3 or a video rip gives the model less to work with than a WAV or FLAC file.

A workflow that holds up

1. Start from the best source you have, a WAV, FLAC or high-bitrate file rather than a ripped stream.
2. Use the highest-quality setting, even when it is slower; both independent tests did.
3. Try two tools on the same song. MusicTech’s conclusion was that results vary by track and that using more than one service can help.
4. Check the stems on headphones, soloed, before building on them.
5. Clean up by hand where it matters, with a spectral editor such as SpectraLayers or iZotope RX.

Is it legal to remove vocals from a song?

Removing vocals from a song for your own practice or karaoke at home is different in law from publishing or selling the result, and the vendors’ own terms reflect that. This is not legal advice, and the rules differ by country.

Separating a song creates a derivative of the recording, not a new right. The stems still contain the original performance and composition, so releasing a remix, sampling an acapella or uploading a cover backing track generally needs permission from the rights holders. Moises states that extracting stems from streaming links would be “an unauthorized derivative work” and that users should process only music they hold the rights to (Moises). LALAL.AI’s terms (updated 30 July 2026) make users “solely responsible for the usage and distribution of uploaded audio files”.

Your own recordings, and properly licensed material, are the safe cases. Moises states that users own their outputs and can use them commercially when they hold the rights to the input (Moises). Remix contests and sample packs that supply official stems avoid the problem entirely.

Frequently asked questions

What is the best AI vocal remover?

LALAL.AI is the best AI vocal remover for most people, because it is the only browser tool near the top of both independent 2026 tests: third of 11 in MusicRadar’s January 2026 test, behind two desktop applications, and the standout on drums in MusicTech’s July 2026 test. For practising on a phone, Moises is the better app, with 4.7 from about 28,000 App Store ratings. Apple Logic Pro’s Stem Splitter won MusicRadar’s test outright but needs a Mac with Apple silicon and costs $199.99.

What is the best free AI vocal remover?

MVSEP is the best free online vocal remover, because it allows 50 separations a day and MusicTech’s July 2026 test rated its vocals the best of nine tools. Ultimate Vocal Remover is the best free option that runs on your own computer, with no limits and no uploads, and MusicRadar called its MDX-Net vocals “exceptional and entirely lossless”. Moises’s free plan allows five uploads a month and LALAL.AI’s free tier allows 10 minutes.

What is the best AI stem splitter?

Apple Logic Pro’s Stem Splitter scored highest in MusicRadar’s January 2026 test of 11 tools, with 16 out of 20, and splits six stems: vocals, drums, bass, guitar, piano and other. Outside the Mac, LALAL.AI is the strongest browser-based stem splitter, placing third in the same test and standing out on drums in MusicTech’s test. The one independent human rating of commercial services, in Meta’s December 2025 paper, scored AudioShake, Demucs, Moises and LALAL.AI within 0.10 of each other on professional material.

Is LALAL.AI better than Moises?

LALAL.AI is better for removing vocals, and Moises is better for practising music. In MusicRadar’s January 2026 test LALAL.AI placed third with 15 out of 20, while Moises missed the top five because of “warbling, clicks and spill” in its vocal stems. Meta’s December 2025 human ratings put the two within 0.04 of each other on professional material. Moises has far more user ratings, a free allowance of five uploads a month and practice tools such as tempo and key change.

Can AI remove vocals from any song?

AI can separate vocals from almost any mixed recording, but no tool is lossless on every song. Reverb, backing vocals and instruments in the same range as the voice leak between stems in every tool tested by MusicRadar and MusicTech in 2026. Results also vary by track, so testers recommend trying more than one tool on the same song and starting from a high-quality file such as WAV or FLAC.

Is it legal to use an AI vocal remover?

Using a vocal remover on your own recordings, or on a song for private practice, is generally low risk, but publishing or selling stems from a song you do not own usually needs permission from the rights holders, because the stems still contain the original recording. Moises refuses streaming links on the grounds that extracting stems from them would be an unauthorised derivative work, and LALAL.AI’s terms make users solely responsible for how they use uploaded audio. Rules differ by country, and this is not legal advice.

Does Logic Pro have a vocal remover?

Yes. Logic Pro’s Stem Splitter separates a recording into vocals, drums, bass, guitar, piano and other, and has done since Logic Pro 11.2 on 28 May 2025. It runs on the Mac without uploading the file and needs a Mac with Apple silicon. Logic Pro costs $199.99 as a one-off purchase, and its Stem Splitter won MusicRadar’s January 2026 test of 11 stem separation tools.

Is Ultimate Vocal Remover safe and still updated?

Ultimate Vocal Remover is open-source software under the MIT licence, published on GitHub by its developer Anjok07, and it processes files on your own computer without uploading them. Its last stable release, version 5.6, dates from 2023 according to its releases page, and its newest Roformer and SCNet models are available only in a beta build. Download it from the project’s GitHub releases page rather than third-party sites.

What does SDR mean for vocal removers?

SDR, or signal-to-distortion ratio, measures in decibels how closely a separated stem matches the original isolated recording, so a higher number is better. The best academic models score about 12 to 13 dB on vocals on the MUSDB18-HQ test set, led by ByteDance’s BS-RoFormer at 12.72 dB in September 2023. SDR does not always match what listeners prefer: expert listeners in the 2023 Sound Demixing Challenge rated the top three systems a statistical tie, and the organisers found no clear correlation between human judgement and SDR.

What is the difference between a vocal remover and an AI music generator?

A vocal remover separates an existing recording into its parts, such as vocals and instrumental, without creating anything new. An AI music generator such as Suno creates a new song from a prompt. Some generators, including Suno, can export stems of songs they made, but that is a different job from separating a song recorded elsewhere. Our page on the best AI music generators covers the generation side.

How to choose

1. For most people, start with LALAL.AI. Its free 10 minutes are enough to test it on your song before paying $7.50 a month.
2. If you are practising an instrument or singing, use Moises for its tempo, pitch and chord tools and five free uploads a month.
3. If you already use Logic Pro, use its built-in Stem Splitter before paying for anything else.
4. If you want to pay nothing, use MVSEP online or Ultimate Vocal Remover on your own computer.
5. If you need to clean the result by hand, use SpectraLayers Pro.
6. Whatever you pick, try two tools on the same song, start from the best-quality file you have, and check you have the rights before you publish anything you separated.