The full data tables are on the research hub. This post summarises the key findings.
Short answer: HTDemucs (base) achieved the highest mean SDR at 8.38 dB across all four stems on the test set. It averaged 1.8 seconds per 7-second track on M4.
The speed range across models was notable: HTDemucs 6S (6-stem) processed tracks in 1.6s on average, while HTDemucs FT (fine-tuned) took 6.4s. For most producers the quality difference justifies the time cost, but batch processing scenarios change that calculation.
Full model comparison: /research/model-comparison/
Yes, measurably. MP3 128kbps input produces stems with a mean SDR of 7.8 dB, compared to 8.04 dB for 24-bit WAV. That’s a 0.24 dB difference. The vocal stem shows the most degradation, which makes sense given how MP3 compression handles mid-frequency content.
For most production use, MP3 320kbps is close enough to lossless that the quality difference is small. Below 192kbps the degradation becomes more consistent.
Full format comparison: /research/format-quality/
In 4-stem mode, summing the separated stems back produces a reconstruction that correlates with the original at r = 0.996. The difference signal sits at -21.6 dB below the original, meaning there’s real but limited information loss in the separation process.
6-stem separation shows higher reconstruction error, which is expected: dividing the signal into more components means more rounding and leakage at each split.
Full reconstruction data: /research/reconstruction-fidelity/
The strongest correlate of vocal SDR across the test tracks was Chroma variance (harmonic complexity) (Pearson r = 0.522). In plain terms: tracks where the harmonic content is clearly separated from percussive content in the frequency domain tend to separate more cleanly. Dense arrangements with a lot of frequency overlap between instruments are harder for the model to disentangle.
Full analysis: /research/complexity-prediction/
All five tests, methodology notes, and full data tables are on the research hub. Tests ran on March 14, 2026 using MUSDB18-7s (the 7-second sample of the standard MUSDB18 benchmark dataset) on Apple M4 with MPS acceleration. I’ll re-run and update these when HTDemucs releases a major model update.
]]>The underlying technology for both tools sits in similar territory, built on architectures like HTDemucs and MDX-Net that represent the current state of AI-based source separation. On a typical well-produced pop or electronic track, either tool will give you a usable stem set. Upload, process, download. That’s the workflow on both platforms, and neither requires any technical knowledge to operate.
The quality gap between these tools and older approaches like early Spleeter is real and audible. Bleed and artifacts still exist, but they’re much less common than they were even a few years ago, and both tools sit near the top of what browser-based separation currently offers.
LALAL.AI has a strong reputation for vocal isolation specifically. The vocal stem on vocal-heavy pop and R&B material is genuinely clean, and that’s not marketing, it’s something you can verify by running the same track through both tools. For DJs or producers who are primarily pulling vocals from songs, LALAL.AI is a credible choice.
StemSplit.io is stronger on multi-stem separation. When you need clean drums, a usable bass stem, and separated melodic elements rather than just a vocal/instrumental split, its 4-stem and instrument-specific outputs hold up well. Producers working on arrangements or sampling work tend to find more value there.
LALAL.AI’s vocal quality is a real competitive strength, and it’s worth being direct about that. StemSplit.io isn’t behind it by a large margin, but if vocal isolation is your only need and the material is vocal-heavy pop, LALAL.AI is the stronger choice. StemSplit.io’s advantage shows up more clearly when you need the full breakdown.
LALAL.AI offers vocal/instrument separation as its base, with some higher-tier plans unlocking additional stem types. StemSplit.io offers full 4-stem separation (vocals, drums, bass, other) and more granular instrument-level options on paid plans. For producers who need more than a vocal and an instrumental, the breadth of options at StemSplit.io is a practical advantage.
If you’re comparing the two on a specific task like making a karaoke track or isolating a vocal for sampling, both do it. Where they differ is in what you can do beyond that.
Both tools are fast. Upload times depend on your connection and file size, but for a typical 3-4 minute track, processing completes quickly on either platform. StemSplit.io tends to be slightly faster for typical track lengths, though this varies enough that it’s not a deciding factor. You’re not waiting around with either service.
LALAL.AI uses a credit-based system. You buy credits and spend them based on minutes of audio processed. This makes sense for occasional or irregular users who want to pay only for what they use, but it adds mental overhead when you’re trying to budget a project with a lot of tracks.
StemSplit.io offers a free tier with some limitations (no batch processing on free, for example) and a subscription for heavier use. The subscription model is simpler to reason about if you’re processing tracks on a regular basis. There’s no mental math on how many minutes a batch of stems is going to cost.
For a more thorough breakdown of free and paid options across the broader landscape, this post on free vs paid stem splitters goes into more detail.
If your primary use case is clean vocal isolation from vocal-heavy material, LALAL.AI is worth trying first. It also suits users who process audio at irregular volumes and prefer paying per use rather than committing to a monthly fee. Some commercial users prefer the metered model specifically because it maps cleanly to per-project cost tracking.
Producers who regularly need full multi-stem breakdowns, not just vocals, will find StemSplit.io’s output and workflow better suited to the work. Music educators who need to break down arrangements for students, beatmakers who want to sample specific instruments, and anyone processing enough tracks to benefit from a flat monthly cost are natural fits.
StemSplit.io’s free tier is functional for casual use, though if you’re processing a high volume of tracks, you’ll hit its limits. That’s a real constraint worth knowing going in.
Both tools are building on similar underlying model architectures. The specific tuning and implementation matters, but neither is doing something categorically different from the other at the core level. If you’re curious about how browser-based tools compare to running models locally, the online vs desktop comparison covers that tradeoff in detail.
For most producers, the real question is simple: do you need great vocals, or do you need great everything? LALAL.AI wins the first question more often than not. StemSplit.io wins the second.
]]>That’s the pitch, anyway. The reality is a little more nuanced, which is why this guide exists.
This is the complete picture: what stem splitting is, where it came from, how the AI actually works, what you can realistically expect from it, and how to choose the right tool for what you’re trying to do. If you’ve landed here without much background, start reading straight through. If you already know the basics, use the sections below to jump to what you need.
A stem, in audio production, is a bounced mix of a subset of tracks from a session. A producer might send a vocalist a “stem” containing just the drums and bass, so they can record over it. A mixing engineer receives stems (drums, synths, vocals, FX) rather than hundreds of individual tracks. The word predates AI by decades.
What’s changed is the direction of travel. Traditional stems always flowed outward: you had the session, you bounced the stems, you shared them. AI stem splitting reverses this. You start with a finished stereo mix and work backward, trying to reconstruct what the individual elements sounded like before they were mixed together.
That reconstruction is genuinely difficult. When you mix drums, bass, vocals and guitars into a stereo file, you’re collapsing dozens of tracks into 2 channels. Frequencies overlap. Reverb from the snare smears into the vocal range. The kick and the bass guitar share almost the same frequency band. You can’t perfectly “un-mix” a track any more than you can un-bake a cake.
What AI does is make a very educated guess, and it turns out those guesses are often good enough to be genuinely useful.
This is different from remixing an isolated vocal because someone posted the a cappella online. It’s different from having access to the original multitrack stems from a record label. Stem splitting works on any song, from any era, with no cooperation required from whoever made it.
The idea of separating mixed audio into its sources is called Music Source Separation (MSS), and researchers have been working on it for a long time, see Wikipedia’s overview of stem separation for the academic history.
Early approaches, dating back to the 1990s and 2000s, relied on signal processing tricks: matrix stereo decoding, phase cancellation, non-negative matrix factorization. These worked poorly and were narrowly applicable. The “karaoke mode” on old amplifiers that removed center-panned vocals? That’s a phase cancellation trick, and it’s why the bass guitar often disappeared too.
The research community organized formal evaluation campaigns (SiSEC, the Signal Separation Evaluation Campaign) from around 2008 onward, which gave researchers a standardized benchmark for comparing their methods. Progress was slow.
Then deep learning arrived.
Around 2017, neural network approaches started dramatically outperforming everything before them. By training on multi-track studio recordings (specifically the MUSDB18 dataset, which contains 150 songs with separate stems), models learned what a kick drum “sounds like” in frequency terms versus what a bass guitar sounds like, and could start disentangling them. The research published by Meta AI’s team on Demucs represents one of the most significant milestones: a model that processes audio end-to-end using a waveform-based encoder-decoder architecture, later combined with spectrogram processing in HTDemucs.
The practical result was that quality crossed a threshold where the outputs were actually usable for music production purposes, not just as research curiosities.
The short version: the AI converts your audio into a spectrogram (a visual representation of frequency vs. time), then learns to produce a “mask” that separates each instrument’s contribution from the whole.
Your audio file, which looks like a waveform, actually contains thousands of overlapping frequencies happening simultaneously. A spectrogram makes those frequencies visible as a 2D image: time runs left to right, frequency runs bottom to top, and brightness shows how loud each frequency is at each moment. The kick drum appears as a cluster of bright pixels in the low-frequency region at each beat. The vocal appears as a shifting pattern in the mid-frequency range.
The neural network learns to identify those patterns from having seen (well, heard) thousands of examples of isolated instruments. During training, it sees both the mixed spectrogram and the target (e.g., “what just the drums look like”), and learns to predict a mask that, when applied to the mix, leaves only the drums behind.
The full technical breakdown goes deeper on spectrograms, time-frequency masking, and why some genres are much harder to separate than others. But the key intuition is: this is pattern recognition applied to sound.
Demucs (specifically HTDemucs, the current state of the art) uses a hybrid approach that processes both the raw waveform and the spectrogram simultaneously, combining information from both domains. That’s why it tends to outperform older models on complex material.
Quality is measured using Signal-to-Distortion Ratio (SDR), where a higher number means less of the other instruments are bleeding into your target stem.
Standard stem splitting produces 4 outputs:
The “other” stem is a catch-all, and for dense arrangements it can be a muddy mix of things that don’t naturally belong together. This is one reason why going beyond 4 stems is sometimes valuable.
Some tools and models now go further, offering 4-stem vs 6-stem separation where guitar and piano are separated from the “other” bucket into their own tracks. Whether that extra granularity is useful depends heavily on what you’re doing with the stems.
The files come out as WAV or FLAC, usually at the same sample rate and bit depth as your input. If you put in a 44.1kHz/16-bit MP3 (which has already lost information from compression), you’ll get 44.1kHz/16-bit WAVs. The AI doesn’t recover information that was discarded during MP3 encoding.
Each stem is full stereo, matching the original track’s length exactly. You can import them directly into a DAW and the timing will be perfect.
Here’s where honesty matters.
Stem splitting works best on music that was recorded and produced in a studio with clean separation between instruments in the frequency domain. A polished pop track, a well-produced hip-hop beat, a simple singer-songwriter recording: these tend to separate cleanly.
Dense orchestral music is genuinely hard. The violin section, the cellos, the brass, and the woodwinds all overlap massively in frequency space, and the AI struggles to disentangle them. Live recordings with room bleed are harder than studio recordings. Genres with a lot of reverb (shoegaze, lo-fi) are harder than dry productions.
You will get artifacts. This is not a flaw unique to one tool, it’s inherent to the problem. Bleed and artifacts are what you call it when pieces of one instrument show up faintly in another stem. The hi-hat might ghost into the vocal stem. A bit of bass might leak into the drums. How much this matters depends on what you’re doing with the output.
For karaoke tracks and practice purposes, minor bleed is irrelevant. For professional remixing work where you need pristine isolation, it may be a dealbreaker unless you spend time cleaning up the stems manually in your DAW.
The models available (Demucs, MDX-Net, HTDemucs) have meaningfully different strengths on different material. Using the best model for your specific genre and use case makes a real difference.
The range of applications is wider than most people expect when they first encounter this technology.
Remixing and production. The most obvious use. Grab the vocal from a released track and build a new instrumental around it. Or isolate the drum pattern from a song you love and study exactly how the groove was constructed. Sampling and beatmaking workflows have been transformed by the ability to cleanly (or at least cleanly enough) isolate individual elements without having to chop around them in a sample.
More specifically, see stem splitting for sampling and beatmaking for how producers actually integrate this into their workflow.
Karaoke. The simplest application: remove the vocals, keep everything else. Making a karaoke track is one of the most common reasons people try stem splitting for the first time. The quality is usually good enough for home use, especially on modern pop where the vocal is well-separated in the mix.
Learning songs by ear. Isolate just the bass to figure out a bassline. Pull out just the guitar to learn a chord progression without the vocal distracting you. Using stem splitting to learn by ear is one of the most genuinely underrated use cases, especially for instrumentalists.
DJ performance. Modern DJ tools have started integrating stem separation directly into their software so DJs can do real-time vocal swaps, drum replacements and instrumental blends between tracks. Stem splitting for DJs covers this in detail, including which DJ software has it built in.
Audio restoration. Less talked about but increasingly useful: separating stems to clean up problematic recordings, reduce noise, or rescue audio where one element is obscuring another. Stem separation for audio restoration covers the specific workflows involved.
Practicing with real tracks. Vocalists slowing down a song to learn it, horn players figuring out a solo, pianists trying to hear a comping part underneath a busy arrangement. The isolation that stem splitting provides makes this far easier than trying to pick out an instrument from a full mix.
There are three main axes along which stem splitting tools differ: online vs desktop, free vs paid, and standalone vs DAW-native.
Online tools run in a browser, require no installation, and work on any computer. The tradeoff is that you’re uploading your audio to someone’s servers, which matters if you’re working with unreleased music. Processing speed depends on server load. Online vs desktop stem splitters covers the full comparison.
Desktop tools run locally, which means your audio never leaves your machine. They’re generally faster once set up (especially with a good GPU), allow batch processing, and give you more control over which model you use. UVR5 (Ultimate Vocal Remover) is the most flexible free desktop option, supporting Demucs, MDX-Net and several other models.
Free vs paid is a real distinction. Free tools work, but paid tools often offer better models, faster processing, cleaner UX, and features like stem refinement and noise reduction. The free vs paid breakdown walks through exactly what you get at each price point.
DAW-native integration is becoming more common. Ableton Live 12’s built-in stem separation lets you separate a clip directly in the arrangement view without leaving the DAW. Logic Pro has had a vocal separation feature for several versions. These integrations are convenient, but the quality doesn’t always match dedicated tools.
For most people who want to try this without installing anything, StemSplit.io is the online tool to start with. It uses HTDemucs under the hood, processes quickly, and the interface is straightforward. Drag in a file, get 4 stems back.
For the full breakdown of what to use when, see online vs desktop stem splitters and free vs paid stem splitters.
Once you’ve got your stems, you need to do something with them.
Bringing stems into your DAW covers the practical steps for importing, aligning and working with stem files in Ableton, Logic, FL Studio and other DAWs. There are a few gotchas with sample rate matching and track alignment that are worth knowing before you start.
If you specifically want to isolate vocals, see how to isolate vocals from a song. If drums are what you’re after, how to extract drum stems goes into the specific considerations for percussion isolation including model choice.
This guide is meant to be an entry point. Every section above has a corresponding deep-dive post:
Understanding the basics:
Choosing the right setup:
Specific use cases:
Have a specific question? The Stem Splitter FAQ covers the most common ones.
]]>It’s easy to assume the choice is just about convenience, but there’s more to it than that. Online tools and desktop software often run different underlying models, with different optimization priorities. An online service might default to a model that handles pop and rock vocals cleanly but struggles with jazz. A desktop tool lets you choose the model yourself and swap it out if the first one doesn’t perform well on your source material.
The separation model matters, and understanding the models that underpin these tools gives you a realistic sense of what to expect from either approach. Meta AI’s Demucs repository is where the open-source model development happens, and reviewing what different versions were trained on makes the output quality differences less mysterious.
No installation is the obvious one, but it matters more than it sounds. Desktop stem separation software typically requires a reasonably modern GPU to run efficiently. If you’re on a lower-spec machine, a MacBook Air without a discrete GPU, or a laptop you use mostly for things other than audio, the online approach sidesteps the hardware problem entirely.
Online tools also tend to update silently. When the underlying model improves, you get the improvement automatically without doing anything. Desktop software requires you to stay on top of updates, sometimes reinstall components, and occasionally deal with breaking changes.
The accessibility angle is real too. A music teacher setting up ear training exercises for students doesn’t need everyone to install and configure software. An online tool means you send a link, not a setup guide.
Batch processing is the clearest case. If you’re working through an album catalog, pulling stems from 50 tracks for a DJ set, or processing a sample library, uploading tracks one at a time to a web interface is genuinely painful. Desktop tools like UVR5 let you queue hundreds of files and walk away.
Desktop tools also give you the ability to pick your model directly. You’re not relying on whatever the service has configured as default. If HTDemucs fine-tuned performs better on your specific source material than the standard version, you can use it. That level of control simply isn’t available in most online tools.
Upload size limits don’t apply when you’re processing locally. A 45-minute live recording in WAV format can exceed 500MB, that’s a problem for most online tiers but not for desktop software.
Processing speed also tends to be faster on desktop if you have a GPU, assuming the software can use it properly.
This is genuinely overlooked. If you’re uploading an unreleased track to a third-party web server, that audio exists on someone else’s infrastructure. Most reputable services have privacy policies that address this clearly, but reading those policies before uploading an album you haven’t released yet is reasonable. Not paranoid, just careful.
For producers working under NDAs, handling client material, or developing tracks they plan to commercially release, the question of where the audio goes isn’t trivial. Desktop software processes everything locally. Nothing leaves your machine. That’s a real differentiator for some workflows, not a marketing point.
In practice, most people who do any serious volume of stem splitting end up using both. Online for quick jobs, one track, testing whether separation is clean enough on a specific song, making a karaoke version on the fly. Desktop for batch work, when quality really matters, or when the source material is something they’d rather not upload.
This isn’t a compromise, it’s just matching the tool to the task. The Complete Guide to Stem Splitting goes deeper on the broader landscape if you want context on where everything fits.
The free vs paid question overlaps with this too. UVR5 desktop is free with a learning curve; good online tools have a cleaner experience but usually cost something past a certain usage level.
Start online. The reason is simple: you find out quickly whether stem splitting does what you need for your specific use case, without installing anything or spending time on configuration.
StemSplit.io is the right starting point for online use. No setup, no installation, results in under a minute for most tracks. The interface doesn’t get in the way, which matters when you’re trying to evaluate whether the output quality is good enough for your use case, not learn a new piece of software.
If you hit real limitations, queue times, batch processing needs, upload size restrictions, or privacy requirements, that’s when desktop software becomes worth the setup. UVR5 is the desktop tool most serious users end up with. It’s not beginner-friendly, but it’s the most capable free option available.
The Stem Splitter FAQ covers a lot of the common questions about both approaches, including what file formats different tools accept and what output quality to expect. And if you’re thinking about how stems fit into a DAW workflow, that post has practical guidance on what to do with the stems once you have them.
You don’t need to choose a side. Use the online tool until it stops doing what you need, then look at desktop. That’s the actual path most people take, and it works.
]]>That said, there are real situations where paying makes sense. Here’s an honest breakdown.
Most free online stem splitters put constraints somewhere in the stack. The common ones:
Not every service imposes all of these. Some are more generous with free access than others. But if you’ve tried a free stem splitter and been frustrated, one of these constraints is usually the reason.
For a lot of use cases, free is completely fine.
You want to isolate the vocal from one song to practice singing over it? Free. You’re making a karaoke track for a single event? Free. You’re a student learning a song by ear and want to isolate the guitar part to transcribe it? Free.
The output quality from free tiers on mainstream songs is usually pretty clean. Models trained on pop and rock music handle those genres well, and the free tier often uses the same model as paid, just with processing constraints layered on top. Testing before you commit to a subscription is also genuinely useful; run a track through free, hear what the stems sound like, and decide if the quality is good enough for what you need.
Professional workflows are where the math changes.
If you’re a producer working on 10 tracks a week and you need stems for each one, the queue wait and usage caps of a free tier become a real cost in time. Paid plans typically offer unlimited (or much higher) monthly processing, priority queuing, and batch upload. That’s not a luxury for a working professional; it’s just necessary.
Output quality on paid tiers can be meaningfully better, particularly for complex material. Dense orchestral arrangements, jazz recordings with a lot of harmonic overlap, lo-fi recordings with noise floor issues: these are situations where the better model makes an audible difference. Progress in stem separation research is real, but the gap between an older model and a current one is most visible on exactly this kind of difficult material. For vocal isolation from a straightforward pop track, you might not notice. For a live jazz recording with multiple soloists, you probably will.
Commercial licensing is another real consideration. Free tiers sometimes have terms that limit commercial use of the output. If you’re using stems in a professional context, track, sync license, or live performance, the paid tier usually comes with clearer rights.
Ultimate Vocal Remover (UVR5) is free desktop software and it’s genuinely powerful. It bundles access to open-source models including Demucs from Meta AI Research and various MDX-Net variants. You can run multiple models, switch between them, tune settings, and process locally without upload limits. For someone who wants maximum control and quality without paying a subscription, it’s the real answer.
The trade-off is the learning curve. UVR5 is not a simple drag-and-drop interface. It requires installation, some understanding of the different models available, and willingness to experiment with settings. If you’re comfortable in that environment, it’s excellent. If you just want to paste in a link and get stems back in 60 seconds, it’s not the right tool.
This is a different kind of free: free as in you do the work of setting it up. That’s a valid choice, it’s just a different value proposition than an online free tier.
For online use, StemSplit.io is where to start. The quality is high, the interface is straightforward, and you don’t need to install anything or understand model architecture to get a good result. It’s accessible enough for casual use but capable enough that professionals are using it too.
The free access lets you test it before committing, which is exactly the right way to evaluate any audio tool. If you hit the limits of free and find yourself frustrated by queue times or processing caps, the paid upgrade is priced reasonably relative to the time it saves.
If you’re doing high-volume batch work and you’re comfortable with desktop software, UVR5 plus your own processing setup is worth the investment. It’s the more powerful option in absolute terms. But for most musicians, producers, and educators, the tradeoffs involved in running desktop software aren’t worth it, and a good online tool does the job well.
The Complete Guide to Stem Splitting covers the full landscape if you want broader context. If you’re deciding between online and desktop approaches, there’s also a dedicated breakdown in Online vs Desktop Stem Splitters. And the Stem Splitter FAQ addresses a lot of the common questions about both free and paid options.
Pay when the free tier is actually limiting your work. Don’t pay to feel like you have a professional tool. Those are different reasons, and only one of them is worth money.
]]>The phrase gets used loosely to describe at least 3 different situations:
Isolating a problematic element. You have a mix where one instrument sounds bad, and you want to pull it out to process it separately, then maybe reconstruct the mix. This is the most legitimate use of stem splitting in a restoration-adjacent context.
Cleaning up the full mix. Someone hands you a messy recording and asks if stem splitting can fix it. Usually the answer is no, at least not directly. Splitting a bad mix into stems just gives you bad stems.
Cleaning an individual stem after extraction. You split the audio, then apply noise reduction or other processing to the individual stem. This can work, but it introduces its own artifacts from the separation step before you’ve even started the restoration work.
Understanding which scenario you’re actually in matters a lot. The approach that helps in one situation actively makes things worse in another.
There are real use cases, and they’re worth knowing.
The clearest one: you have a live recording with significant crowd noise or room ambience mixed in with the music. Running it through a model like HTDemucs to pull out the music from the noise floor can give you a cleaner music signal to work with. The separation won’t be perfect, but it can give you something more workable than the original, especially if the musical content is reasonably prominent in the mix.
Another legitimate use: you recorded something, the mix is genuinely bad, but the performance is worth saving. Splitting the stems lets you work with individual elements, correct levels, maybe re-record something over the isolated instrumental, and rebuild from there. This isn’t restoration in the traditional sense, it’s more like reconstruction. But the result can be significantly better than trying to fix a bad mix with EQ and dynamics alone.
There’s also the archival scenario: you have an old recording where one element, say a lead vocal from a live performance, needs to be isolated for documentation or preservation. The stem splitter won’t give you a studio-clean vocal, but it can give you something more intelligible than the original source, which matters for archiving historical performances. This connects to how these models actually work under the hood, which is covered in How AI Stem Separation Actually Works.
Clipping is the big one. If your source audio is clipping, running it through a stem splitter doesn’t fix the clipping. It passes the clipped signal through a neural network that was trained on clean audio and has no idea what to do with it. You’ll get separated stems that still clip, plus whatever artifacts the model introduced trying to make sense of a distorted waveform.
Similarly, if you’re dealing with severe phase issues, heavy saturation baked into a mix, or significant low-frequency distortion, stem splitting won’t clean any of that up. The model is separating sources based on learned patterns of what instruments sound like. Corrupted audio doesn’t match those patterns cleanly, so separation quality degrades, sometimes badly.
There’s also an artifact compounding problem. Stem splitting introduces its own artifacts: bleed between stems, frequency smearing at the edges of the separation. If you then apply noise reduction on top of that, you’re stacking 2 lossy processes. The artifacts and bleed post goes into detail on what separation actually does to audio quality. Worth reading before you use this approach on anything you care about.
A persistent hum from an amplifier that’s only affecting one instrument in a recording. If you can separate that instrument into its own stem, you can apply targeted hum removal to just that stem without touching the rest of the audio. This is genuinely useful and one of the cleaner applications of this workflow.
Isolating a vocal from a lo-fi live recording for archiving. You won’t get a clean studio vocal, but if the goal is to create a document of what was performed, even a voice-forward stem is more useful than the full mix for that purpose.
Extracting a clean enough melody from a rough demo to hand off to a session musician for re-recording. The stem doesn’t need to be broadcast-quality; it just needs to be clear enough to communicate the part. That’s a lower bar, and stem splitters often meet it.
The model you use matters more in restoration-adjacent work than in standard stem splitting. HTDemucs and MDX-Net models have different strengths, and some are better at handling difficult source material than others. Meta AI’s Demucs repository documents what different model versions were optimized for, which is worth checking if you’re choosing a model for a specific restoration task.
If your audio has significant noise floor problems, clicks, pops, crackle, severe room reverb, or any kind of codec damage, you need dedicated restoration software. iZotope RX is the industry standard and handles these problems directly. Spectral repair tools in RX can isolate and remove specific problem frequencies without touching the rest of the signal. That’s fundamentally different from what stem splitting does.
Stem splitting works by learning what instruments sound like and separating them. Restoration tools work by identifying unwanted artifacts and removing them. These are different problems with different solutions, and there’s not much crossover.
That said, combining both workflows does sometimes make sense. Run restoration on the source file first to remove noise floor and clicks, then use stem splitting on the cleaned audio via a tool like StemSplit.io. In that order, you’re giving the separation model cleaner input, which produces better output. Doing it the other way around (split first, then restore each stem) generally produces worse results because you’re working with audio that’s already been degraded by the separation process.
The Complete Guide to Stem Splitting covers the fundamentals if you want a broader picture of what these tools are and aren’t designed to do. For restoration-specific questions, the Stem Splitter FAQ addresses some of the common misconceptions about what separation can fix.
The honest bottom line: stem separation is useful in a specific slice of restoration-adjacent work. It’s not a restoration tool, it’s a separation tool, and that distinction matters. Use it where it actually helps, reach for RX or similar when you’re dealing with real damage.
]]>Take out the bass and drums from a complex chord progression and suddenly the voicings become obvious. That’s the thing most musicians don’t realize until they try it: your ears are constantly prioritizing. The low end grabs attention, the rhythm locks you in, and the harmonic content gets processed almost as background.
When you strip a track down to just keys or just guitar, you start hearing things you’d genuinely never caught before. Suspended chords that resolve differently than you assumed. A piano part doubling the melody an octave up. A rhythm guitar playing a slightly different pattern than the lead. These aren’t things you could miss through carelessness; they’re things the mix hides by design.
This is especially true for dense productions. Pop records in particular are built so that every element occupies its own sonic space, which means individually, each part sounds thinner than you’d expect. Isolating them reveals the actual part, not the part as your brain stitched it together from the full mix.
What you pull out of a stem splitter depends on what you’re trying to learn.
Guitarists usually benefit most from a stem without the bass. The low-end muddiness between bass guitar and rhythm guitar is real, and when they’re separated, chord shapes become much clearer. You can hear whether a chord is open or barred, whether there’s palm muting, whether the strumming pattern is straight or syncopated.
Drummers learning a complex fill want the drum stem in isolation. Full stop. Listening to drums in a dense mix means your brain is constantly filtering. Pull the drum stem and you hear every ghost note, every hi-hat variation, every subtle shift in the kick pattern. If you’re studying a particular drummer’s style, the isolated stem is like getting a transcript.
Pianists and keyboardists probably get the most out of 6-stem separation when it’s available. The standard 4-stem split combines all instruments into “other,” which can lump together piano, organ, strings, and synths. A 6-stem model that isolates piano specifically is genuinely useful for transcription. (More on stem model types in The Complete Guide to Stem Splitting.)
Vocalists have their own specific use case covered below.
The workflow is pretty simple once you have your stems. Run the song through a stem splitter like StemSplit.io, download the individual tracks, and load them into whatever you use to play audio. Most people just use a DAW, but even a basic audio player works.
A few things that help:
This is something that takes a little while to notice, but it’s real: isolating instruments doesn’t just teach you a song, it teaches you how parts relate to each other.
When you isolate a bass line and play it alongside the isolated chord stem, you start hearing things like how the bass outlines chord tones on the downbeat, or how it creates tension by sitting on a non-chord tone against the harmony. That’s music theory made audible. It’s the kind of thing you can read about in any textbook, but hearing it in a real recording makes it click differently.
The same applies to rhythm. Pulling out just the drums and just the bass from a funk track, then playing them together, makes the groove relationship between those two instruments obvious in a way that no amount of description can replicate. According to Wikipedia’s overview of stem separation, this kind of analytical listening has applications well beyond practice, including musicology and education. The practical application for musicians is straightforward: use it.
Here’s the use case that doesn’t get talked about enough: using the instrumental stem for vocal practice.
If you isolate the vocals from a song, the byproduct is always an instrumental version with the vocals removed. That instrumental is surprisingly useful for singers. You can practice your part while hearing the full arrangement, without the original singer in your ear competing with you. It’s more useful than a karaoke track in some ways, because it’s the actual recording, just without the vocal.
This works well for audition prep, for singers learning harmonies, and for vocal coaches running lessons. Load the instrumental into any audio player, sing along, and you get direct feedback on your intonation against real instrumentation. Combined with a simple recording setup so you can hear yourself back, it’s a legitimate practice tool. You can also check out how to make a karaoke track for a slightly more polished version of this process.
For drummers specifically, there’s also a dedicated post on extracting drum stems if you want to go deeper on that particular use case.
Stem splitting for music practice isn’t a replacement for traditional ear training. You still need to do the work of internalizing what you’re hearing, identifying intervals, building musical memory. But it removes the biggest obstacle: not being able to clearly hear what you’re trying to learn. The Stem Splitter FAQ has more on what formats and models are available if you’re just getting started.
StemSplit.io is a good first stop if you want to try this without installing anything. Upload a track, get the stems, and start a practice session. The interface is simple enough that you’re not spending time figuring out the tool instead of learning the music.
]]>Here’s how producers are actually using it.
The classic approach to sampling involves flipping a section of the original recording, a bar or two of drums, a chord hit, a bassline fragment. The craft is in what you choose and how you chop it. But you’re always working with the full mix. If you want the guitar from a song where the guitar is buried under a vocal and three keyboards, you’re stuck with all of it.
Stem splitting changes the raw material. Pull the guitar out as its own track, and now you’re working with something close to a direct recording of that performance. The texture is there, the playing style is there, but the congestion of the original mix isn’t.
The Meta AI Demucs repository is where most of the separation technology underlying these tools originates. The academic work on source separation (see the HTDemucs paper on arXiv) has moved fast enough that what’s available in consumer tools today would have seemed unrealistic just a few years ago.
Not every stem is usable for sampling, and knowing what to listen for saves time.
Drum stems from well-produced modern records tend to be clean. The kick and snare are usually well-separated by the model because they’re spectrally distinct from vocals and instruments. Drum bleed from other elements is the thing to watch for, a bit of low-end bass note bleeding into the kick pattern, or room reverb from a piano sitting under the snare. How to extract drum stems covers what to realistically expect from the drum stem specifically.
Basslines are hit or miss. A clean, isolated electric bass against a sparse production will separate well. A bass synth that shares frequency range with a pad or keyboard will blend with it in the model’s output. Listen before you commit to building around it.
Chord stabs, piano hits, and guitar parts work well when the original arrangement is sparse. Dense productions with lots of harmonic overlap are harder for the model to untangle. A stab that sits alone in the mix for even a bar is often enough to get something usable.
The general rule: the more separation there is in the original production between elements, the better the stems will be.
These are genuinely different source materials and they produce different results.
A chop is a slice of the full mix. You’re grabbing a moment in time where everything in the production sounds the way you want it: the kick, the bass, the melody, the room. The texture is rich and self-contained because it contains everything. That’s the soul of classic boom-bap sampling.
A stem sample is an isolated element. It has more flexibility (you can pitch it, filter it, layer it) but it lacks the density of the full mix. It often sounds thinner on its own, which isn’t a problem if you’re treating it as one piece of a larger construction, but it’s a different creative tool than a chop.
Neither is better. They’re different approaches that suit different styles. A lot of producers end up using both: a drum stem for the rhythmic foundation, then chopping melodic elements from the full mix for a richer texture.
The most practical use of stem separation in beatmaking is treating the isolated stem as one layer in a mix you’re building from scratch.
Pull the piano stem from a soul record. Loop a bar of it. Now build a drum pattern from your own drum samples underneath it. Add a 808 bassline you’re playing yourself. What you’ve made is mostly original, with one element sourced from the record. This is how producers have always worked, the stem just gives you cleaner access to that one element.
The stem works as a texture, a starting point, or a structural piece. What surrounds it is yours. The more original material you add, the more the finished beat sounds like a production rather than a flip.
Processing the stem before building around it helps too. A bit of saturation or tape simulation can make a clean digital stem feel more analog and connected to the drum sounds you’re laying under it. Light filtering to remove frequencies you’re not using keeps it from fighting with your 808 or kick.
There are 2 ways producers typically use stems in a larger workflow.
The first: start from a stem and build out. You’ve found an element you like, a drum pattern or a chord sequence, and you build the whole track around it. The stem is the seed. This is the traditional flip approach, just with more control over which element you’re flipping.
The second: build an original production first, then use stem separation when you need a specific texture you can’t create synthetically. You’re mostly writing your own material but you want a specific live guitar feel, or a particular kind of snare sound. You find a recording that has what you need, split the stem, extract it, and drop it into your arrangement.
Bringing stems into your DAW covers the technical side of importing and organizing stems once you have them. Getting the file format and alignment right from the start makes the rest of the process smoother.
If you end up with a stem that has quality issues, clicks, or artifacts from the separation process, AI stem separation for audio restoration has useful techniques for cleaning up problem material before you build around it.
The legal side of stem sampling is basically the same as the legal side of regular sampling: stems from copyrighted recordings are still copyrighted, and isolating an element doesn’t change that.
For personal production and private listening, using stems is generally fine. Nobody’s coming after you for making beats in your bedroom with isolated drums from a record.
For commercial use, meaning anything you release, sell, or monetize, the stem carries the same clearance requirement as a chop of the full mix. If you’re releasing a track commercially that contains an isolated drum hit from a copyrighted recording, you need clearance (or you need the sample to be unrecognizable enough to constitute interpolation, which is its own legal gray area). The fact that you split the stem yourself doesn’t affect the copyright status of the underlying recording.
The FAQ covers the legal questions in more detail. The short version: personal use is fine, commercial use without clearance is the same risk it’s always been with sampling. Stem splitting doesn’t create a new legal category.
StemSplit.io is a fast option for getting stems to experiment with. Upload a track, get the separated elements back in a couple of minutes, and start finding out what’s usable. The complete guide to stem splitting is worth reading too if you want a fuller picture of how the separation technology works and what to expect from different kinds of source material.
]]>But there’s a real gap between what stem splitting promises and what it delivers in a live DJ context right now. Worth being honest about both sides.
For most of DJ history, your options with a track were: play the whole thing, EQ out some frequencies, or blend at the cue point. You were working with the full mix as a single unit.
Acapellas changed that slightly, but official acapellas exist for a tiny fraction of released music. Stem splitting effectively creates an acapella (and an instrumental) from any song that exists. That’s a significant shift in what a DJ can do with their library.
The Wikipedia overview of stem separation has good context on how the field developed from academic research into something practical enough for consumer tools. What was PhD-level research 10 years ago is now a 2-minute web upload.
A producer working on a remix might want 6 stems or more: vocals, drums, bass, piano, guitar, synths. The more granular the separation, the more control they have.
DJs usually don’t need that level of detail. In practice, most DJ use cases come down to 2 outputs: the acapella and the instrumental. Sometimes 4: vocals, drums, bass, and everything else. The “other” or “accompaniment” stem is often all you need for a DJ context, because you’re blending full tracks, not deconstructing them note by note.
This means you don’t need to overthink the stem count when prepping for a gig. A 4-stem split covers almost everything a DJ needs. 4-stem vs 6-stem separation gets into the differences if you want the longer version.
Real-time stem splitting in a live DJ set is mostly a future promise.
The latency introduced by current separation models makes live processing impractical for most performance contexts. The models need time to analyze audio before they can output the separated components, and even the fastest implementations introduce delays that are audible and disruptive in a club environment where timing is everything. Native Instruments has experimented with real-time stem features, and Pioneer has included some stem-adjacent tools in their ecosystem, but none of it yet performs well enough for confident live use in front of an audience.
The practical approach right now: prepare your stems before the gig. Split the tracks you plan to use, export the acapellas and instrumentals you need, and load them into your DJ software as separate files.
This is actually a better workflow anyway. You know exactly what quality you’re getting. There are no surprises mid-set.
Why stem splitters aren’t perfect covers the artifacts to listen for when you’re evaluating your pre-split stems. Knowing what bleed sounds like helps you make better decisions about which tracks to use.
With a library of pre-split stems, the creative possibilities are real and immediate.
Using an acapella from one artist over a completely different instrumental is the most obvious move, and it works well when you get the key and tempo right. Most DJ software lets you key-shift stems without changing tempo, so a semitone or two of adjustment is easy.
Extending breakdowns is another strong application. If you have the drum stem from a track isolated, you can loop it for as long as you want and bring the rest of the elements back in on your own timeline. It gives you control over tension and release that you don’t get from the standard track structure.
Layering bass elements from one track over the drums of another is more technical but effective when it works. The low-end bleed issue (more on that below) is something to screen for here, since you don’t want two competing kick drums if you’re using the bass stem from a different track.
Bringing stems into a DAW is relevant if you want to do any pre-processing before loading stems into your DJ software. A bit of light EQ on a stem before the gig can clean up issues that would be annoying to deal with live.
For pre-gig stem preparation, web-based tools are the fastest option. Upload the track, get the stems back, download and add to your library. StemSplit.io works well for this: the turnaround is fast, you get WAV stems, and there’s no software to install or update.
For integrating stems into your DJ workflow, most modern software handles audio file import fine. Rekordbox lets you load audio files and treats them like any other track. Serato handles imported audio the same way. The stem just appears as a track in your library, you cue it up like anything else.
The UVR5 tool (available on GitHub) is worth knowing about if you’re doing high-volume stem preparation and want to run things locally on your own machine. It’s more technical to set up, but it’s fast and free once it’s running.
Bleed is the main issue, and it’s more audible in a loud club environment than it is on headphones.
Kick drum bleed into the bass stem is the most common problem. If you’re using an isolated bass stem alongside a different kick drum, you may hear ghost kicks from the original track underneath your chosen drums. In a quiet room it’s subtle, in a loud room on a big sound system it can be obvious.
Vocal bleed into the instrumental is the other one to listen for. Most separation models handle lead vocals well, but backing vocals, doubled lines, and vocal chops that sit deep in the mix sometimes survive into the instrumental stem partially. You’ll hear a ghostly presence. Whether that matters depends on how you’re using the stem.
The honest take: pre-listen to every stem you plan to use before you play it. Not just a quick scan, actually listen through at full volume with headphones designed for critical listening. What you miss on a quick check is what will catch you out during a set.
For background on how the AI separation actually works and why these artifacts show up, the complete guide to stem splitting is a solid starting point. And the FAQ has answers to some of the common questions around using stems in DJ contexts, including questions about licensing and commercial use. For fast pre-gig stem preparation, StemSplit.io is the quickest option.
]]>This isn’t about which DAW you use. The workflow is the same in Logic, Ableton, FL Studio, or anything else.
File format is the first decision, and it’s an easy one: always download stems as WAV, not MP3, if you have the option.
MP3 encoding introduces artifacts. When you’ve got 4 separate MP3 stems and you’re processing them individually, those artifacts can become audible, especially in the high frequencies and on transients. WAV files don’t have this problem. A 24-bit WAV is what you want for any production work. 16-bit is fine for practice or reference listening, but 24-bit gives you more headroom when you’re applying EQ and compression.
If your stem splitting tool only offers MP3 downloads, that’s workable, just know you’re starting with a compressed source. For anything you’re putting into a real release, it’s worth using a tool that exports WAV. StemSplit.io exports WAV stems by default.
The other thing to check before you even open your DAW: the sample rate of your stems. Most modern tools output at 44.1kHz. If your session is set to 48kHz, your DAW will either resample on import (introducing subtle quality loss) or play back at the wrong speed. Match your session sample rate to your stems, or vice versa, before you start.
Create a new session and set the tempo to match the original track before you import anything.
This matters more than people think. If your session tempo is wrong by even a few BPM, the stems will drift out of sync with any loops, MIDI patterns, or additional recordings you add. There are a few ways to find the tempo: use a tap-tempo tool, run the track through a BPM analyzer, or use your DAW’s tempo detection (Ableton’s Analyze function is good for this, and Ableton’s own stem separation docs cover how they handle this natively in Live 12). Apple has a similar walkthrough in the Logic Pro stem splitter documentation if that’s your setup.
Set the project tempo, then create a stereo audio track and import the original full mix to use as a reference. Keep it muted once you’ve confirmed the session length. You’ll want it there to compare against.
Import all 4 stems (or however many you downloaded) at the same time. In every major DAW, you can multi-select audio files and import them together, which places them starting from bar 1. Do not import them one at a time and drag them in separately, because the starting point needs to be identical for all of them.
Once they’re in, name the tracks clearly. “Vocal,” “Drums,” “Bass,” “Other” is clear enough. If you’ve done a 6-stem split and have more granular tracks like “Guitar” or “Piano,” name them specifically. Color coding helps too: most producers use blue for vocals, green for bass, orange or red for drums. Pick a system and stick to it.
A quick visual check: all stems should have exactly the same length as the original track. If one stem is even a few milliseconds shorter, something went wrong on the export. Re-download before you build anything around it.
Stems produced by a splitter are in phase with each other. That’s by design: the models are trained to output components that sum back to something close to the original when you layer them together.
The thing to verify is that all stems start at exactly the same point in your timeline. If you dragged one stem in after the others and it’s 10 milliseconds late, summing the stems won’t give you what you expect. At the sample level, even a few samples of offset causes comb filtering. It’s subtle but audible, especially on the low end.
To check: zoom all the way into bar 1 in your DAW and confirm every stem’s waveform starts at the same position on the timeline. They should all line up to the exact sample. If they don’t, nudge them back into alignment.
What are audio stems? covers the underlying concepts if you want a clearer picture of why phase coherence matters when working with separated tracks.
Treat stems as raw material, not finished tracks.
The vocal stem will almost certainly need some processing before it sits right in a mix. The drum stem may have low-end bleed from the bass, especially on the kick. The bass stem can have thump from the kick bleed going the other way. These are properties of how source separation works, not signs that you did something wrong. How AI stem separation actually works explains why this bleed is essentially unavoidable with current models.
Light high-pass filtering on everything except bass and drums helps clean up unwanted low-end. A gentle low-pass on the bass stem can reduce kick bleed without affecting the fundamental. On the vocal stem, a bit of de-essing and a gentle reverb tail removal can help it sit more naturally.
The key word is light. These stems have already been through one stage of processing. Heavy-handed EQ and compression on top of that starts to feel like over-cooked audio.
Once everything is imported and aligned, the fun part begins.
Swapping the drum stem is one of the most effective moves. Mute the original drum stem and drop in a different loop or sample at the same tempo. The vocal and bass stems continue playing as they were, but the rhythmic foundation is entirely new. This is a fast way to reinterpret a track.
Muting the bass stem and replacing it with an original bassline gives you the most creative control over the low end. You keep the drums and the feel of the original, but you’re writing the bass part yourself. That’s a genuinely useful technique for remixing and beatmaking.
Using the vocal stem from one track and the instrumental elements from another is where things get interesting. It requires careful key and tempo matching, but a well-chosen combination can produce something that sounds entirely original.
If you’re building a DJ set around pre-split stems, the same import principles apply, you just end up with a library of individual elements rather than a single project.
For drum-focused work, how to extract drum stems covers what to expect from the drum stem specifically, including how much bleed to expect from kick and room mic artifacts.
The complete guide to stem splitting has a broader overview of the whole process if you’re early in figuring out how this fits into your workflow. The DAW side of things rewards a bit of upfront organization, it’s faster to set up properly the first time than to sort out alignment and naming issues later.
]]>