Riffloop Review 2026

When I put together the roundup of the best AI stem separation tools in 2026, almost every engine in it shared one assumption: you give it a file, and you get audio you can take away. Riffloop breaks the assumption at both ends. There is no file going in, because it works on the YouTube video already open in front of you, and there is no file coming out, because it will not give you one.

That second half is the thing to understand before anything else, and it is the reason Riffloop was never a candidate for the numbered list in that article. It is not a worse separator than the eight engines there. It is a different category of software that happens to contain a separator, and if you evaluate it as a separator you will conclude it is broken. Judged as what it is — a practice instrument for people learning parts off records — it does something none of those eight do.

Written by YECK, founder of MixingGPT. Every product fact below was read out of Riffloop’s public documentation on 8 September 2026 and is linked at the point it is used, so you can check any of it in about a minute. Where the marketing copy and the help pages disagree, both versions are quoted rather than reconciled — those gaps are the most useful thing I found.

Quick Overview: Nine Riffloop Claims at a Glance

If you want the thirty-second version, here it is. Full details are below the table.

Stem Separation Has Quietly Split Into Two Products

The engines got good enough that the interesting question stopped being how cleanly a model can pull a vocal out and became what you are allowed to do with the result. Two answers have separated out, and most confusion about tools in this space comes from treating them as one market.

Separators produce assets. Demucs through UVR, Lalal.ai, Logic’s Stem Splitter, RipX, SpectraLayers, Audioshake. You hand them audio you have the right to process, and they hand back files at a known sample rate that open in a DAW. The file is the product, and everything about how they are judged — artefacts, bleed, bit depth, export formats — follows from that.

Practice tools produce a performance you can play along with. Moises sits partly here, and Riffloop sits entirely here. The separation is a means to an end, and the end is that the guitar is gone so you can be the guitar. Nothing needs to persist. Once you can play the part, the separated audio has served its purpose and can evaporate.

Riffloop takes the second position further than anything else I have used, to the point of designing the file out entirely. Two of its help pages carry the same warning in nearly identical words — it is a practice tool, not a downloader, and nothing is saved to your computer as an audio file — and both then repeat it as a flat “no” in their FAQs. Putting that in the help pages rather than burying it in terms of service is a deliberate signal, and it is the whole architecture of the product in one sentence.

It also explains the constraint that shapes everything else: because streamed audio is never retained, a section has to be captured in real time while the video plays. The Studio guide lists what breaks a capture — pausing, seeking, changing the speed, switching to another video, or stopping the tab share — and a ninety-second chorus takes ninety seconds to capture. Cloud separators beat that comfortably, and they beat it because they are allowed to hold the file.

There is one accommodation worth crediting, because it is the sort of thing that only gets built after someone has been annoyed by the alternative: a Use what I’ve got control that stops a capture early and keeps whatever has been recorded so far, rather than discarding the attempt and the credit with it.

The Nine Claims, Checked One at a Time

These are the claims Riffloop makes on its homepage, in its help pages, and in the pitch its founder sent me. Each gets the claim as stated, what the documentation actually supports, and whether that is the same thing.

1. Six Stems on Faders, Under the Player, With Nothing to Upload

The claim: install the extension, open any YouTube song, and a panel appears under the player offering a six-fader mixer with no upload step.

What is actually there: the panel offers four modes — loop a section, remove vocals, build a backing track, and full stem separation — and the last of those produces the mixer. The six stems are vocals, drums, bass, piano, guitar, and solo and lead. Each row has a level percentage, an S to solo it and an M to mute it, with soloing dropping the other five to zero. Four presets set all six at once, and Riffloop publishes the exact fader values behind each: instrumental is the vocal at 0% with everything else at 100%, rhythm only leaves drums and bass, vocals only is the inverse.

Two details are better than they need to be. The waveform carries most replayed chips pulled from YouTube’s own data, which usually land on the chorus or the solo — often exactly the section you were going to select anyway, and a genuinely smart use of metadata that is already on the page. And separation keeps running if you switch tabs, with a notification when it finishes.

Where it falls short: the sixth stem has three different names across Riffloop’s own pages. All three help pages call it Solo & Lead. The homepage free-plan feature list calls it Other — “All 6 stems — Vocals, Drums, Bass, Guitar, Piano, Other” — and the homepage stem section calls it “and the rest.” Those are not the same promise. “Other” is the residue bucket every four-stem model produces, and “solo and lead” implies a model that identifies a lead line as a musical role. The help pages are the more specific and more recently detailed of the two, and the Studio does mark a detected solo in the mixer, so the capability appears real. The homepage is the page that needs correcting.

2. It Runs on Your Device and Nothing Is Uploaded

The claim: Riffloop’s homepage states “on your device — nothing uploaded” as a headline property of the product.

What is actually true: of the extension, this holds, and the free tier description repeats it as “runs locally in the Extension — nothing uploaded.” The extension’s permissions are also narrower than most: it reads and changes data on youtube.com and studio.riffloop.app and can show notifications, and Riffloop states it does not run anywhere else. For a tool that has to inject UI into a page and tap its audio, that is a tight scope.

The Studio is a different story, and the homepage headline does not distinguish them. The Studio accepts uploads up to 50 MB, requires an account, and offers two separation qualities where the better one is a Pro feature. A local model that produces two quality tiers gated by subscription is unusual; a server-side model that does is routine. Riffloop has not published where Studio separation runs, and the honest summary is that the extension’s local claim is specific and credible while the site-wide version of it is doing more work than it should.

The homepage category-comparison table makes the tension visible in a single column. It gives Riffloop a tick for “works on your own uploaded files” and a tick for “runs on your device — nothing uploaded” four rows apart. Both are true of the product as a whole and neither is true of the same surface at the same time, which is exactly the kind of claim that needs splitting in two.

Why it matters: if privacy is the reason you are considering this, stay in the extension and do not assume the guarantee travels with you into the Studio. For unreleased material of your own, a local separator such as Demucs through UVR remains the option with nothing to reason about at all.

3. Karaoke Leaves the Vocal In at 15% Rather Than Stripping It

The claim: the karaoke preset is not a full vocal removal. It leaves the original vocal at 15% as a guide.

What is actually true: exactly this, published as a number in a preset table. Karaoke is the vocal at 15% with the other five stems at 100%; instrumental is the vocal at 0%. Riffloop separates the two presets explicitly and explains why, which is that a quiet guide vocal tells you where the line sits while you sing over it.

This is the detail that most convinced me the product was built by someone who practises. A 15% vocal is a musical decision, not a technical one, and it is the opposite of what a spec sheet rewards — most tools in this space compete on how completely the vocal disappears. It also quietly solves the artefact problem: at 15% the residue in the instrumental stems reads as part of a deliberately present guide vocal rather than as separation failure. Whether that was the intent or a happy consequence, it works in the product’s favour.

Where it falls short: 15% is fixed. There is no documented way to set it to 5% or 30%, and the natural next step once you know the line is to fade the guide down. You can get there manually by leaving karaoke and pulling the vocal fader yourself, so nothing is lost, but the preset is a one-way door.

4. Slow a Song Down to 25% Speed With the Pitch Unmoved

The claim: Riffloop’s homepage advertises a minimum playback speed of 25%, twice — once in the hero copy as “down to 25% speed” and once as a headline statistic, “min speed 25%.”

What the documentation says: 0.5×. The extension guide lists the speed row as 0.5×, 0.75×, 1×, 1.25× and 1.5×. The Studio guide lists the same five steps and states that they are the same on both surfaces. Nothing in any of the three help pages I read describes a 0.25× setting existing anywhere in the product.

I cannot tell you which is wrong, only that they cannot both be right. It may be an unshipped feature that reached the homepage early, or a documentation gap. Either way, plan on 0.5× being the floor, and treat the 25% figure as unconfirmed until Riffloop says otherwise.

Why the gap matters more than it looks: the difference between half speed and quarter speed is the difference between practising a fast passage and transcribing one. At 0.5× a sixteenth-note run at 140 bpm is still moving at 70 bpm, which is not slow enough to hear individual note attacks in a legato line. That is precisely when a guitarist reaches for the speed control. Half speed covers most practice; quarter speed is for the cases where practice has already failed, and it is the reason dedicated slow-down and transcription software has historically gone further than the halfway mark. Five fixed steps with no setting below 0.5× is the single most restrictive spec in the product.

The pitch-preservation half of the claim is not in doubt. Riffloop states it repeatedly across all three help pages, and the chord documentation adds a detail that supports it: chord positions follow the song’s actual position rather than its playback speed, so slowing down does not desynchronise the chart.

5. Transpose the Song Across a Full Two Octaves

The claim: shift the whole track up or down by up to twelve semitones in either direction, with the tempo unaffected.

What is actually true: as documented, on both the extension and the Studio, with the offset shown in the section header so you always know how far you have moved. Two octaves of range is generous. It covers bringing a song into a comfortable vocal register, matching a capo position, and the case that catches people out — a record cut at a non-standard reference pitch, or a guitar tuned a half step down for the whole album.

The Studio adds something the extension does not: it detects and displays the key, along with tempo and time signature, so you can see what you are transposing from rather than working it out. The extension detects tempo only, for the click.

Where it falls short: the honest caveat on any transposition of a finished record is that large shifts cost audio quality, and separated stems have less margin than an untouched mix. Riffloop makes no claim about how far you can go before it degrades, and there is no published guidance on it. Two or three semitones is routine anywhere; the outer reaches of a twelve-semitone range are a different proposition, and the range being available is not the same as it being usable throughout.

6. A–B Loop Any Section of the Song

The claim: drag a region over the part that trips you up and let it cycle.

What is actually true: two different things, depending on whether the audio has been separated. Plain looping on untouched audio is unrestricted and free, and spends no credit, because nothing is being separated. Looping a separated section is capped: one minute on the free plan in the extension, seven minutes on Pro, and no cap in the Studio.

One minute is enough for a chorus, a verse or most solos, and Riffloop’s own tip is that tighter selections separate more cleanly, so the limit pushes you toward better results. It is not enough to work through a whole song with the stems up, which is the case where the seven-minute Pro allowance starts to matter.

The best detail here is that a saved loop keeps your stem balance with it, so it returns the way you left it rather than as a bare time range. Both surfaces save loops, and signing in syncs them between the extension and the web. The searchable library — filtering by label, song or stem, with Your songs and Your sections views — is documented for the Studio; the extension is listed as saving loops without that browsing layer described.

Where it falls short: the homepage promises that loop regions “snap to the beat,” but none of the three help pages describe beat snapping — they describe dragging markers and now buttons that drop a marker at the current playback position. Nor does anything say how the loop behaves at the seam. Whether the return is sample-accurate or takes an audible gap is what decides if a looper survives an hour of repetition, and it is unaddressed.

7. Backing-Track Mode Shows Live Chord Diagrams and a Click Locked to the Song

The claim: pick the instrument you want to play yourself, and Riffloop mutes it and hands you the key, the chords as fretboard shapes, and a beat-synced click.

What is actually there: the most complete feature in the product, and the part with the least competition. According to the backing-track guide, one job returns the detected key, a chord strip naming the chord at every point in the song, any solo sections found with timestamps, the same six-stem mixer with your chosen instrument at 0%, and a metronome with a count-in. Clicking through to the chord shapes gives five diagrams at once — the current chord enlarged in the middle, the previous and next either side — advancing with the song as it plays, with a toggle between beginner open-position shapes and barre chords.

Several design decisions here are better than they had to be. The chords come with every backing-track job regardless of which instrument you muted, rather than being a separate paid step. The five-diagram window solves the real problem with chord charts, which is that seeing only the current chord gives you no time to prepare the next shape. The easy-or-barre toggle acknowledges that the same chord has different right answers depending on how long you have been playing. And solo detection identifies whichever instrument is actually carrying the solo in that passage rather than assuming it is the one you muted, so a track with a piano break gets labelled as one.

Where it falls short: no accuracy claim is published for either chord or key detection, and both are genuinely hard problems. Chord recognition on dense modern production, borrowed chords, slash chords and inversions is where these systems usually show their limits, and there is no stated confidence indicator. Riffloop also does not say whether the fretboard shapes cover instruments other than guitar, and the documentation follows a guitar example throughout. Bass and keyboard players should assume the chord names apply and the diagrams may not.

8. Free to Try, With No Signup

The claim: free to try, no signup, no card.

What is actually true: the no-signup part is real and unusual, and it deserves credit. You install the extension, open a song, and separate it without creating an account. Almost nothing in this category lets you do that.

The allowance behind it is smaller than “free to try” suggests, and Riffloop’s pages give three different numbers for it:

Two pages say three total and one says three per day, which is a material difference: one is a trial and the other is a usable free tier. The conservative reading, and the one stated most emphatically, is three separations total as a guest.

The mechanic underneath is clear and fairly explained. Every separating action costs one credit, and stem separation, vocal removal and backing tracks all cost the same, from a pool shared between the extension and the Studio. Only plain looping is free, because nothing is separated. Practically, three guest credits means one careless attempt on the wrong section is a third of your trial, and the section length cap on free is one minute. Sign in before you start.

9. Export Is Available in the Studio

The claim: as Riffloop’s founder put it to me, export is available on the Studio.

What is actually true: export exists for audio you uploaded yourself, and for nothing else. Riffloop’s own Studio comparison table is unambiguous: the extension gets “No, on any plan,” and the Studio gets “Files you upload only. MP3, M4A or WAV,” with WAV reserved for Pro. The same page states directly that a track streamed from YouTube or SoundCloud can be practised with but not exported, and its FAQ answers the question outright: to export stems as files, upload your own audio.

So the export answer for the use case Riffloop is built and marketed around — the song already playing on YouTube — is no, on every plan, on both surfaces, permanently. This is the single most important fact in this review, and it is the one most likely to be lost in a one-line summary.

What you do get, on an upload: more than the comparison table implies. You can export individual channels, all channels at once, or an export mix that prints your current fader balance as a single file — which is a genuinely useful option, because it means the karaoke or rhythm-only balance you built can leave as audio. The metronome click at the detected tempo is exportable as its own channel too. Two limits attach. WAV is Pro, so the free tier is MP3 or M4A. And free-plan exports include only the first 60 seconds, which is stated on the Studio page but appears nowhere in the plan comparison on the pricing page.

One inconsistency worth flagging rather than concluding from: the Studio page lists the exportable individual channels as vocals, drums, bass, piano, guitar and the click, and does not name Solo & Lead among them — even though it is one of the six stems everywhere else on the same page. That is more likely an incomplete list than a missing feature, but if the lead line is the stem you need out, confirm it before you pay.

The honest read: this is a design decision, not an oversight, and it is almost certainly what lets the product exist. An extension that hands you audio files extracted from YouTube is a downloader wearing a practice-tool badge, and downloaders have a short and eventful history. By never producing a file from streamed audio, Riffloop stays a playback tool. I would rather use a product that made that choice deliberately than one that had it forced on it later.

But the shorthand needs correcting, because “export available on the Studio” is precisely the phrase a reader in a hurry would take to mean their YouTube stems can come out somewhere. They cannot. If you own the file already, the Studio will separate it and give you the stems back, and that is a real feature — it is just a different product from the one in the demo video.

Extension or Studio? They Are Not the Same Product

Riffloop is two surfaces sharing one engine and one credit pool, and which one you should use is decided almost entirely by where your audio comes from. Both give you the same six stems, the same four presets, the same five speed steps and the same twenty-four semitones of transposition. Everything else differs.

Rows reproduced as worded in Riffloop’s Studio documentation, read 8 September 2026. The guest figure is the one that conflicts with two other Riffloop pages, so it is quoted rather than reconciled.

The short version: use the extension when the song is on YouTube and you want to play along, and the Studio when the file is yours, when you need longer than seven minutes, or when you want to keep the stems. One friction point worth knowing is that streamed tracks in the Studio still have to be captured in real time, and Chrome will ask you to share the tab with audio enabled. If tab audio is off the capture is wasted and the credit is spent.

Pricing, Plan by Plan, and What Each One Actually Unlocks

Riffloop has three price points and a free tier. What follows is what each one changes in practice rather than what the pricing page says, because the interesting part is which limit you hit first.

Free — $0

Six stems, all four presets, the full speed and transpose range, loops up to one minute, saved loops, and local processing in the extension. Three separations as a guest, five a day once you sign in. No card. The limits that bite are the one-minute section cap in the extension and the fact that you are on the standard separation engine rather than the better one; if you are uploading your own files to the Studio, exports are also trimmed to the first 60 seconds. As a way to find out whether the workflow suits you it is genuinely sufficient — you can learn a chorus on the free tier and nothing is crippled.

Pro — $5.95/month or $39/year

Five things change, in descending order of how much they matter. Sections go from one minute to seven in the extension, and the Studio drops its length limit entirely, which is the difference between practising a part and working through a song. Separations become unlimited, so you stop rationing. You get the Best quality engine instead of standard. Exports stop being trimmed to the first 60 seconds and WAV unlocks, both for uploaded files only. And saved loops sync across your devices.

The annual price is the one to take: $39 a year against $5.95 a month works out at about $3.25 a month, which is the 45% saving Riffloop advertises. For context, Moises runs from free to around $10 a month, so Riffloop’s annual plan costs roughly what four months of that does. At this price the comparison is not really with other software anyway. It is with the twenty minutes you would otherwise spend scrubbing a YouTube video back to the same four bars.

Founder’s Lifetime — $99 once

Everything in Pro, permanently, including future Pro features, described on the site as a limited-time launch offer.

The honest read on pricing: $99 against $39 a year breaks even at about two and a half years, so this is not a discount, it is a bet that Riffloop is still running and still shipping in 2029. Lifetime tiers at launch usually exist because early revenue is worth more to a young company than the future subscription it gives up, which is a reasonable trade for both sides but not a risk-free one. If you practise several times a week and the workflow has already earned a place in your routine, the maths is fine. If you are not sure yet, the annual plan costs less than half as much and tells you the same thing. I would take the year first.

One structural note on the free tier, in Riffloop’s favour: the credit that a separation costs is the same whether you use stem separation, vocal removal or a backing track, and plain looping never costs anything. There is no tier where a feature you were shown is withheld until you pay — you get all of it, less often. That is a more honest free plan than most.

10. How It Holds Up on a Busy Midrange

What can be said in advance is what to listen for, because it is the same everywhere in this category. Separation models are good on modern pop, hip-hop, R&B and electronic records, where the arrangement is sparse, the elements are already well separated in the mix, and the model has heard ten thousand records built the same way. They struggle on dense rock, orchestral material and older mono recordings. A crowded midrange is the hard case specifically because guitars, keys, snare body and vocal fundamentals all occupy the same octave, and a model that has to assign energy to exactly one of six buckets has to guess.

Riffloop is also working under one constraint the file-based engines are not: it captures streamed audio in real time from a lossy YouTube stream rather than processing a file at its original resolution. That is a worse starting signal than a WAV, and it sets a ceiling nothing downstream can lift. For playing along it is very unlikely to matter. It is another reason the comparison with export-grade separators is not like for like.

Where Riffloop Sits Next to the Separation Engines

This is the comparison most readers arrive wanting, and the answer is that only one column really decides it.

The full eight-engine breakdown, including RipX DAW, SpectraLayers 12, Audioshake and iZotope RX 12 Music Rebalance, is in the stem separation roundup.

Moises is the direct comparison, since it is the pick I gave to musicians and students learning parts by ear. The two overlap heavily and diverge on one axis. Moises takes an upload, keeps a library, runs on a phone, and lets you download the stems. Riffloop removes the upload and the wait entirely, runs where the song already is, and adds chord diagrams and a click that Moises does not match for guitarists. Then it refuses to give you a file.

So the decision is not about quality, it is about where your practice happens. If you learn from YouTube on a laptop and never need the audio again, Riffloop has less friction than anything else here, and the chord work is a real advantage. If you want the stems on your phone at rehearsal, or a setlist you keep, or anything that continues after the browser tab closes, Moises is the more complete product and remains my recommendation for that case.

Against the file-based engines there is no contest in either direction, because they are not answering the same question. Nobody choosing between Demucs and Riffloop is confused about quality; they are confused about category.

Who Should Use It, and Who Should Not

Use Riffloop if you learn songs off YouTube and the browser is where your practice happens. Guitarists get the most: the muted-instrument backing track with live fretboard shapes and a beat-locked click is the closest thing here to a lesson, and the easy-or-barre toggle means it stays useful past the beginner stage. Singers get the 15% karaoke guide, which is a better default than a full strip. Bass and keyboard players get the isolation and the chord names, and should not count on the diagrams. Teachers get a way to produce a backing track for a student in about the time the section lasts, with no file management at all.

Do not use Riffloop if your work needs the audio. Remixing, sampling, rescuing a session whose stems are lost, anything that lands in a DAW timeline — it structurally cannot do those with YouTube audio, and using the Studio on your own uploads to get around it means you already had the file, which puts you back in the market the roundup covers. If you need to go slower than half speed, or if you are working with dense orchestral or older mono material where separation gets hard, look elsewhere. And if the appeal is the privacy claim, keep it to the extension.

How This Review Was Done

Every product fact above comes from one of four Riffloop pages, all read on 8 September 2026:

• riffloop.app — marketing claims, the plan comparison, and all pricing.
• How to split a YouTube song into stems — the extension, the six stems, the preset values, speed and transpose ranges, guest credits.
• How to get the chords to any YouTube song — backing-track mode, chord strip, fretboard diagrams, solo detection, credit costs.
• How to use the Riffloop Studio — the extension-versus-Studio table, capture behaviour, export rules and formats, pan, key and time-signature detection.

Where two Riffloop pages disagree, both are quoted and the conflict is named rather than resolved in the product’s favour or against it. Competitor facts — Moises, Lalal.ai, Demucs, Logic Pro Stem Splitter — are carried over from the eight-engine roundup rather than restated from memory, so the two articles cannot contradict each other. Nothing in this review is based on a press kit, a review copy of someone else’s coverage, or a claim I could not find on a page I could link to.

Where This Category Is Going

The separation models are close to done as a competitive axis. When six or seven engines all produce a usable vocal split on a modern record, quality stops being what anyone chooses on, and the differentiator moves to what the software does with the result. That shift has already happened in mixing, which is why the useful tools there stopped being one-click processors and became assistants that explain a decision. It is happening here too, a couple of years behind.

Riffloop is a clear instance of it. Nothing in the product is a separation breakthrough. What is new is the decision to treat the separated audio as disposable and put the effort into the twenty seconds after it exists — the loop, the click, the chord that is coming next. The other consequence of that decision is legal, and it is the more interesting one: a practice tool that never produces a file has a defensible position working on streamed audio that a downloader does not. If this approach spreads, that is why.

The gap it leaves open is judgement. Riffloop tells you what the record is doing. It has nothing to say about what you did when you played it back, which is the harder problem and the one that separates a part you have learned from a part you have learned well. That is the boundary between a practice tool and a mixing tool, and both sides of it are still improving faster than most people are keeping up with.

Frequently Asked Questions

Related Deep Dives