LALAL.AI is a browser, desktop and mobile stem separator from Swiss company Omnisale GmbH that isolates vocals, drums, bass, guitars and six more instrument tracks, with voice cleaning, cloning and an API for developers.
Used this tool? Rate it
Used this tool? Rate it
LALAL.AI is an audio source separation service that takes a finished mix and pulls it apart into its component tracks. The site presents it as an AI vocal remover built for pro-level quality, powered by AI and transformer technology, and that framing is more accurate than most marketing copy: the product does one hard technical thing and builds a family of features around it.
The core operation is stem separation. You upload a song, choose what you want extracted, and receive isolated tracks — vocals on their own, the instrumental without them, or individual instruments split out. Around that core sit six further tools covering voice cleanup, voice transformation and cloning, echo and reverb removal, and separating lead vocals from backing vocals.
The service is operated by Omnisale GmbH, a Swiss company registered at Grundstrasse 10 in Risch-Rotkreuz. That entity name is confirmed in three independent places — the site footer, the Apple App Store publisher field, and the Google Play developer field — which is a stronger identity check than a single About page. The company dates its start to 2020, when it built its first neural network, and reports registered users rising from 3.8 million in 2024 to 6.79 million the following year. That user figure is self-reported and has not been independently audited.
The nearest comparisons are Moises, RipX and the open-source Ultimate Vocal Remover, and each occupies different ground. Moises is built around practice features for musicians — tempo and pitch shifting, chord detection — with separation as one component. Ultimate Vocal Remover is free and runs locally but demands technical setup and hardware. LALAL.AI sits between them: narrower in scope than Moises, considerably easier than a local install, and focused hardest on the separation quality itself.
The most substantive evidence for that focus is the model lineage. The company has shipped a new separation network almost every year: Rocknet in 2020, then Cassiopeia, Phoenix, Orion, Perseus, Andromeda and Lynx in 2026. Seven generations across seven years, with dates corroborated across the official repository, the About page and the changelog, is a verifiable record of sustained investment rather than a marketing adjective.
User sentiment is unusually positive for this category and rests on a large sample. On Trustpilot the service holds 4.3 out of 5 across 4,194 reviews, with 68 percent five-star and 7 percent one-star. That distribution matters as much as the average: it is not the polarised split common among AI subscription tools, and with over 4,000 reviews the sample comfortably exceeds the threshold at which ratings become meaningful.
The company also replies to reviews, including critical ones, which is worth noting because non-response is a frequent complaint pattern elsewhere in this market. What is missing is coverage from major technology press — no independent reporting from TechCrunch, VentureBeat or comparable outlets surfaced during research. The company reports a People's Voice Webby Award for Best Use of AI and Machine Learning, though this could not be independently confirmed in the Webby winners database and the available external coverage traces back to press-release distribution.
The headline capability extracts up to ten stems: vocals, instrumental, drums, bass, guitars, piano, synthesizer, strings and wind. Guitars can be split further into acoustic and electric. This granularity is what distinguishes a serious stem splitter from a simple karaoke vocal remover — extracting a clean isolated bass line from a dense mix is a substantially harder problem than muting a vocal.
Andromeda is the current separation engine, described by the company as its sixth generation. Its rollout is documented rather than vague: Andromeda arrived in November 2025 for vocal isolation and gained drum and bass stem support in March 2026, having extended to voice-and-noise separation in between. That staged expansion tells you something useful — new stem types arrive progressively as the model is trained for them, so capability at any moment is worth checking against the changelog rather than assumed.
Two 2026 additions extend the range. Lynx was trained on a year of hand-cleaned data, while Lyra runs entirely on your hardware with no uploads. Lynx targets voice isolation specifically and is described as smaller and more consistent than its predecessors; Lyra addresses an entirely different constraint, discussed under privacy below.
Voice Cleaner removes background noise and artefacts from spoken recordings, making it applicable to podcast and interview post-production rather than only music. Echo & Reverb Remover addresses room acoustics baked into a recording — a genuinely difficult problem and one of the more useful features for anyone working with imperfectly recorded source material.
Voice Changer and Voice Cloner transform or replicate a voice. On paid tiers these come with Voice Pack Slots, one on Lite and three on Pro, which indicates the feature is properly resourced rather than a token addition. It also carries compliance obligations covered in the limitations section.
The Lead/Back Splitter separates a lead vocal from harmonies and backing layers. Introduced in September 2025 using the Perseus network, this is a niche capability that matters considerably to a specific group — anyone remixing, producing covers, or preparing backing tracks where the harmony stack needs to move independently of the lead.
The web uploader accepts up to 20 files at once in MP3, FLAC, MKV, MP4 and other formats, including video containers. Video input is a practical detail worth noting: you can feed an MP4 directly rather than extracting audio first, which suits dubbing and localisation workflows.
For integration there is a REST API with an OpenAPI specification and an official examples repository, covering separation, voice cleaning, cloning and noise removal. The repository holds roughly 190 stars and documents API V1 as the recommended version. Publicly, however, the API pages do not state pricing, rate limits or quotas — a genuine gap for anyone scoping an integration.
The primary professional application is obtaining stems from material where the multitrack session no longer exists — old recordings, licensed samples, or reference tracks. Producers use isolated drums and bass for remixing, and as a music production tool the VST plugin means this can happen inside the DAW rather than as an export-and-reimport detour.
Removing vocals to produce an instrumental is the most common consumer use, and the reason many people arrive here. The Lead/Back Splitter extends this: you can keep backing harmonies while removing only the lead, which produces a far more usable karaoke track than a full vocal strip.
Voice Cleaner and the reverb remover apply to spoken content as readily as to music. For podcasters working with imperfect room acoustics or noisy field recordings, this is often the more relevant half of the product, and the 4+ age rating and multi-language app support make it accessible to a broad user base.
Separating dialogue from music and effects is a standing requirement in localisation. Direct video file support and the enterprise API make this workable at volume, and the named enterprise customers include dubbing and localisation companies, which corroborates that this is an actual deployed use case rather than a hypothetical one.
Musicians studying arrangements use stem separation to hear how a part sits in a mix — isolating a bassline or a synth layer to transcribe or learn it. This is where the per-instrument granularity earns its value over a simple two-way vocal split.
Companies embedding separation into their own products use the API. The enterprise tier raises the per-file ceiling to 10GB and names customers including Loudly and Dubpro AI, alongside Ollang, Saga, Slate and Voicecheap. Those names are vendor-supplied and were not independently confirmed with the customers themselves.
Paid subscriptions use a two-queue model that is easy to misread. You get unlimited minutes in the Relaxed Queue, with Fast Queue capped at 90 minutes monthly on Lite and 250 on Pro. The Relaxed Queue is not a trial allowance — it genuinely has no monthly cap. What you are buying with Fast Queue minutes is priority processing, not access.
The practical implication: if your work is not deadline-bound, a Lite subscription can process an effectively unlimited volume by accepting slower turnaround. Reserve Fast Queue minutes for jobs where waiting is costly.
Start from the highest-quality source you have. Separation quality is bounded by input quality, and a 128 kbps MP3 will produce noticeably more artefacts than a lossless file. Where possible, feed FLAC or WAV.
Test on a difficult passage rather than a simple one. Dense mixes, heavy compression and reverb-soaked vocals are where models struggle, so evaluating on your hardest material gives a truer picture than a clean, sparse track. Use the free previews for exactly this.
Pay attention to processing order when extracting multiple stems, as user reports indicate the sequence can affect results. If a particular instrument separates poorly, try isolating it directly rather than deriving it from a residual track.
If your material is confidential or unreleased, use Lyra in the desktop app or VST plugin, which keeps processing on your own machine with nothing uploaded. This requires the Pro tier and a desktop install; the mobile apps and web interface do not offer it.
For programmatic use, start from the OpenAPI specification and the official examples repository rather than reverse-engineering requests. Note that API access is included in the one-time minute packs but is Pro-only on the subscription side, so the cheapest route to API access depends on your volume pattern rather than on tier prestige.
This is the single most valuable habit with this tool, and the refund policy makes it essential rather than merely sensible. Free previews let you assess separation quality on your own material at no cost. Once you have purchased and used minutes, refunds are largely unavailable — so the preview is your only real evaluation window.
The two purchasing models suit genuinely different patterns. Subscriptions favour steady ongoing work, especially if you can tolerate the Relaxed Queue. One-time minute packs favour sporadic, project-driven use, since they do not expire monthly the way a subscription allowance resets. Buying the wrong one is the most common way to overspend here.
Because having used any portion of your purchased minutes disqualifies most refund requests, a large pack bought for uncertain future work carries real risk. Buy against committed projects, and consider whether the auto top-up option, which automatically buys the selected package when your balance drops below 150 minutes, actually suits you before enabling it.
The tier differences are not only about volume. The VST plugin, Lyra local processing and API access are Pro-only on subscriptions. If any of these is your actual reason for subscribing, Lite will not do regardless of how many minutes it offers.
Separation models improve — seven generations in seven years is the demonstrated pace. Stems you extract today may be reproducible at higher quality in a year, but only if you have retained the original source file.
The terms are explicit that you are solely responsible for the usage and distribution of uploaded and resulting audio files. Separating a commercial track does not grant any right to use it. This matters most for anyone publishing remixes or covers commercially — the tool creates the stems, it does not clear the rights.
If you use Voice Cloner or Voice Changer in published work, note the synthetic content disclosure obligations under the EU AI Act referenced in the terms. This is a live compliance requirement for European audiences, not a theoretical one.
Producers and remixers needing stems from finished masters. The per-instrument granularity and VST integration make this the practical path when no multitrack exists.
Podcast and spoken-word editors. Voice Cleaner and reverb removal address the most common post-production problems in speech recording, and they work on material that has nothing to do with music.
Localisation and dubbing teams. Direct video input plus API access at volume is precisely the workflow this serves, and existing enterprise customers operate in this space.
Users who need on-device processing. Lyra's local mode is a real differentiator for anyone contractually barred from uploading client material to a cloud service.
Non-English speakers. The mobile app's store listing covers 18 interface languages, which is far better localisation than most tools in this category offer.
Anyone expecting perfect separation. Reviews consistently report audio artifacts, robotic vocal stems, and lingering noises left behind on processed tracks. The technology is good, not flawless, and dense or heavily processed mixes remain hard.
Users wanting practice-oriented features. If you need tempo change, pitch shifting or chord detection alongside separation, Moises is built around those and this is not.
Anyone who might need a refund. All amounts paid are generally non-refundable, including remaining account balances and prepaid fees. Buyers who value flexible refund terms should weigh this carefully.
Free-tier users hoping to produce finished work. The free tier produces previews only and does not allow downloading finished stems. It is an evaluation tool, not a usable free service.
Teams needing certified compliance. The enterprise page claims complete privacy and security but names no SOC 2 or ISO certification, so organisations with formal audit requirements will need to ask directly.
Coverage here is unusually broad. Beyond the web application there are desktop builds for Windows, macOS and Linux, mobile apps on both platforms, and a VST plugin for DAW integration. The desktop installers are downloadable directly from the site.
On iOS, the app holds 4.38 out of 5 across 604 ratings, is rated 4+, and requires iOS 15.0. It was published by Omnisale GmbH in June 2023 and remains actively maintained, updated in August 2026. Notably, its store listing covers 18 interface languages including Chinese, Japanese, Korean, French, German, Russian, Spanish and Portuguese.
On Android, Google Play shows 4.0 stars across roughly 2,670 reviews with more than 500,000 downloads, rated Everyone, updated in August 2026. Reading the three rating sources together gives a useful gradient rather than a single number: Trustpilot 4.3 across 4,194 reviews, iOS 4.38 across 604, and Google Play 4.0 across 2,670. Android scores lowest, which is worth knowing if that is your primary platform.
One version detail is worth flagging: the desktop installer is version 2.18.0 while the iOS build is 2.17.0. Feature parity across platforms is not exact, and Lyra local processing in particular is desktop and VST only.
This is the most commonly misunderstood aspect of the product, so it is worth stating plainly: LALAL.AI sells both subscriptions and one-time minute packs, and both are live. Which is cheaper depends entirely on your usage pattern.
Starter is free, Lite runs $7.50 a month billed as $90 annually, and Pro is $15 a month billed as $180. The site advertises the annual option as saving three months against monthly rates.
Allowances work through the queue system: unlimited minutes in the Relaxed Queue, with Fast Queue capped at 90 minutes monthly on Lite and 250 on Pro. The free Starter tier gets 10 minutes total and no Fast Queue access at all.
Under the Top-Ups tab sit packs that never expire monthly: Master at 750 minutes, Premium at 3,000 minutes for $190, and Enterprise at 5,000 minutes for $300. Master was listed at $50, discounted from $100, at the time of writing — promotional pricing that may not persist.
All three packs include API access, batch upload, stem download, the fast processing queue and a 2GB upload limit. That makes the packs the cheaper route to API access than a Pro subscription for anyone whose volume is bursty rather than continuous.
An auto top-up option automatically buys the selected package when your balance drops below 150 minutes. If funds are insufficient the company sends an email and disables the feature rather than retrying.
Tier differences extend past volume, and these gates decide the plan for many users. The VST plugin, Lyra local processing and API access are Pro-only among subscriptions. Result downloads and batch processing require any paid tier. Upload limits run 200MB per file on the free tier against 2GB on Lite and Pro, rising to 10GB on enterprise agreements.
Read this before purchasing. All amounts paid are generally non-refundable, including remaining account balances and prepaid fees. Refunds are discretionary and limited to unresolved technical faults after 30 days, accidental duplicate purchases with no credits used, and payment processing errors.
Critically, having used any portion of your purchased minutes disqualifies most refund requests. The statutory right of withdrawal requires contact within 14 days and before activating the plan — which means EU consumers who want to preserve withdrawal rights must not start using the service. For subscriptions, cancellation takes effect at the end of the current billing cycle with no refund for unused prepaid minutes.
Moises is the closest competitor and the better choice for practising musicians, bundling tempo and pitch control, chord detection and a metronome around its separation engine. LALAL.AI is narrower and more focused on separation fidelity itself.
Ultimate Vocal Remover (UVR) is free, open source and runs locally with no upload, which makes it attractive for confidential material and for users with capable hardware. The trade-off is setup complexity, model selection by hand, and no support — where LALAL.AI charges for convenience and consistency.
RipX DeepAudio goes further into note-level editing, letting you manipulate individual notes within a separated stem. It is a heavier, more expensive tool aimed at detailed audio repair rather than quick stem extraction.
Audioshake targets the rights-holder and label market with an enterprise focus, and is a closer comparison to LALAL.AI's business tier than to its consumer plans.
Demucs, Meta's open-source separation model, underpins a number of free tools and is worth knowing about as the technical baseline many services build on or benchmark against. Running it yourself trades money for setup effort.
The most consistent criticism across a large review base concerns output quality on difficult material: audio artifacts, robotic vocal stems, and lingering noises left behind on processed tracks. Specific user reports add detail — bass in particular is not always separated cleanly, and the order of separation affects the result. These are inherent limits of current source separation rather than defects unique to this product, but they mean previewing your own material is essential.
Reviewers also flag usability friction: the platform being awkward to navigate, insufficient instructional documentation, and occasional access errors. One user noted difficulty simply finding the download control.
This is the sharpest practical caveat. All amounts paid are generally non-refundable, including remaining account balances and prepaid fees, and having used any portion of your purchased minutes disqualifies most refund requests. For a product where you buy capacity in advance, that combination puts the risk on the buyer. The 14-day statutory withdrawal window closes on activation, not on expiry.
The terms state you are solely responsible for the usage and distribution of uploaded and resulting audio files. Note what is absent: there is no affirmative grant of commercial rights over your output, because the underlying copyright generally belongs to whoever owns the original recording. Separating a track creates stems, not a licence. Anyone monetising the results needs to have cleared rights independently.
Where you use voice cloning or transformation, the terms invoke synthetic content disclosure obligations under the EU AI Act. Neither the cloning consent process nor any abuse-prevention mechanism is documented publicly, and no minimum sample length or voice-authorisation check is described.
The positive is genuinely notable: the policy states the service does not use user files for artificial intelligence training, and does not share copyrighted materials and other uploaded media files with third parties. Both are affirmative commitments, and the second is unusual in this market.
The gaps are equally real. There is no concrete retention period or deletion timeline for uploaded audio — only a reference to a statutory retention period — and no named processor handling audio infrastructure. GDPR rights including erasure without undue delay are set out in full, consistent with the Swiss and EU operating context. For material that genuinely cannot leave your premises, Lyra keeps separation on your own machine, which matters for unreleased or confidential material and sidesteps the question entirely.
The terms do not state a governing law, while the copyright agent sits in Florida and the operator in Switzerland. That split is not explained in the terms, which leaves dispute-resolution venue unclear.
The terms require only that you are of legal age to form a binding contract, with no verification mechanism described. Store ratings are permissive at iOS 4+ and Google Play Everyone. Given that voice cloning is available, the absence of any age or identity check is worth noting even though the content itself is benign.
Several things are simply not published and are therefore not assessed here: API pricing, rate limits and quotas; audio retention duration; the sample requirements and safeguards around voice cloning; and any third-party security certification. Where the company has published nothing, this article says so rather than estimating.
No independent technology press coverage surfaced for this product. The claimed Webby People's Voice award could not be verified in the official winners database, and external coverage of it traces to press-release distribution. The strong evidence here is user reviews at scale and a verifiable engineering record — not press validation.
There is a free Starter tier with 10 minutes and a 200MB per-file limit, but it is an evaluation tool rather than a usable free service: the free tier produces previews only and does not allow downloading finished stems, and batch processing is unavailable. You can judge separation quality on your own material at no cost, which is genuinely useful, but you cannot take finished work away without paying.
Two models run in parallel. Subscriptions: Starter is free, Lite runs $7.50 a month billed as $90 annually, and Pro is $15 a month billed as $180, with annual billing advertised as saving three months. One-time packs: Master at 750 minutes, Premium at 3,000 minutes for $190, and Enterprise at 5,000 minutes for $300. Packs suit sporadic use since they do not reset monthly; subscriptions suit steady use, particularly if you can accept the Relaxed Queue.
It refers to the Relaxed Queue only. Paid subscriptions give unlimited minutes in the Relaxed Queue, with Fast Queue capped at 90 minutes monthly on Lite and 250 on Pro. So volume genuinely is uncapped on paid plans, but priority processing is metered. If your work is not time-critical, this makes Lite considerably better value than the headline minute figure suggests.
Up to ten stems: vocals, instrumental, drums, bass, guitars, piano, synthesizer, strings and wind, with guitars further splittable into acoustic and electric. There is also a Lead/Back Splitter for separating lead vocals from backing harmonies. Available stem types have expanded over time, so the changelog is the authority on what a given engine currently supports.
The tool does not grant you rights it does not hold. The terms state you are solely responsible for the usage and distribution of uploaded and resulting audio files, and there is no affirmative commercial grant, because copyright in the original recording belongs to its owner. If you separated your own recording, your existing rights apply. If you separated someone else's, you need permission from the rights holder regardless of what this tool produced.
No. The privacy policy states the service does not use user files for artificial intelligence training, and also that it does not share copyrighted materials and other uploaded media files with third parties. What the policy does not specify is a concrete retention period or deletion timeline. For material that must not leave your machine at all, the Pro tier's Lyra mode processes locally with no upload.
Usually not. All amounts paid are generally non-refundable, including remaining account balances and prepaid fees, and having used any portion of your purchased minutes disqualifies most refund requests. Refunds are discretionary and limited to unresolved technical faults after 30 days, duplicate purchases with unused credits, and payment errors. EU statutory withdrawal requires contact within 14 days and before activating the plan. Use the free previews to evaluate before buying.
Web, plus desktop builds for Windows, macOS and Linux, mobile apps on both platforms, and a VST plugin. The iOS app holds 4.38 out of 5 across 604 ratings, is rated 4+, and requires iOS 15.0, with 18 interface languages. Google Play shows 4.0 stars across roughly 2,670 reviews with more than 500,000 downloads. Lyra local processing is desktop and VST only.
Better than most, but not flawless. The large review base is positive — 4.3 out of 5 across 4,194 reviews, with 68 percent five-star and 7 percent one-star — yet the recurring criticisms are specific and credible: audio artifacts, robotic vocal stems, and lingering noises left behind on processed tracks, with bass separation singled out as inconsistent. Quality depends heavily on source material, so test a difficult track through the free preview rather than a clean one.
Yes, a REST API with an OpenAPI specification and an official examples repository covering separation, voice cleaning, cloning and noise removal. Access is Pro-only among subscriptions but is included in all three one-time minute packs, so packs can be the cheaper route for intermittent programmatic use. Note that API pricing, rate limits and quotas are not published publicly.
