Humanize Drums: Swing Presets vs AI Randomize Evidence

A swing preset is not humanization. It is a clock delay: every offbeat is shifted by the same fixed percentage, so the feel repeats exactly bar after bar. Human drummers drift; presets don't.

Cubase users can call up humanize through MIDI modifiers and input/output transformers, and DP's Humanize can transform a track from sterile MIDI precision to flawed humanity quickly. But these functions randomize velocity and timing with predictable algorithms—they don't sample from a distribution of human playing. The distinction matters because AI engines like Groove2Groove sample from constrained probability distributions, producing timing variations that resemble a real performance rather than a repeated offset.

Forum comparisons across Logic, Cubase, and Digital Performer focus on ease of use and end result, but none evaluate whether the output passes as human. The evidence from randomization tools shows that true randomness requires a random mechanism—something swing presets lack. The humanize question is whether DAW defaults can match the statistical character of human drumming, or whether AI probability models become the reference.

The Mechanism

Fixed swing presets and AI randomize sound alike on paper but are built on opposite mechanisms. A swing preset is a delay line: in Ableton Live's Groove Pool, the humanize dropdown offers three symbolic swing amounts, and each is implemented by shifting every even subdivision by a fixed percentage of the grid interval. The shift is computed once and repeated exactly, with no variance across bars. A preset is a fixed transformation applied to a grid position, not a model of human timing.

Logic Pro's Humanize MIDI FX and the AI Randomize engine in current DAWs come from the other side of the design space. Both are sampled from a latent-vector model derived from Magenta's Groove2Groove, which was trained on MIDI drum performances (Simon et al.). That training corpus supplies the human component: the model encodes the distribution of onset deviations that actual drummers produced, rather than a percentage lookup. The model does not know a fixed percentage as a setting; it knows a distribution.

At inference time, the engine predicts an onset-offset distribution for every grid position and draws a different offset each time. This is the statistical definition of randomization (Wikipedia): a random mechanism selects a sample from a population. According to Research Randomizer, the service has generated more than 43.7 billion sets of random numbers since 2007, and its example output for a request of five numbers in range 1-50 is 2, 17, 23, 42, 50 — no draw predicts the next. The AI engine's offsets behave the same way: the deviation drawn for an offbeat in one bar carries no information about the deviation for the same grid position in the next bar. Identical input MIDI therefore yields different but statistically similar output takes.

That absence of correlation is the mechanical source of the clock artifact. Because a swing preset shifts an event, not a distribution, a middle setting on a fixed-step pattern moves the same offbeats every bar; the long-short pattern loops as a fixed cycle, so listeners perceive a clock artifact. The AI engine's output has no bar-to-bar correlation, so no such artifact can emerge. This mechanism kills the myth of "swing equals human feel": swing is a musical idiom, not randomness. A live drummer's swing ratio drifts from bar to bar; a preset's does not.

The edge case confirms the mechanism. A fixed swing preset is not wrong because of its amount; it only works when the target is a deliberately stable boombap loop whose swing ratio barely moves, because the human target itself has almost no variance. The moment the groove needs drift, the preset's lack of variance becomes the artifact. DAW users run straight into this confusion: a Cubase user initially used only input and quantize, unaware of humanize entirely (VI-Control forum), and the thread "Comparing humanize functions on different DAWs ..f/b pls" exists precisely because Logic's Humanize behaves differently from equivalent functions in other DAWs (VI-Control forum).

Across the mechanisms, AI Randomize wins for lo-fi and hip-hop: it is the only option that adds cross-bar variance, which is what "humanize" actually means. The middle swing preset is the only fixed delay line worth keeping — and only when a stable boombap groove is the explicit target.

The Evidence

Repp's meta-analysis in Psychonomic Bulletin & Review sets the lower bound. Trained musicians' tap intervals carry a standard deviation that scales with the inter-onset interval. The edge case worth keeping in mind: those are finger-tap data, not full-drum-kit performances. Limb coordination and physical striking add noise, so a real drummer should land at or above that band, never below it. That makes the band a deliberately generous floor for what counts as human timing.

Senn et al. in Music Perception set the ceiling. In isolated drum loops, large random microtiming deviations received lower groove ratings. The word "isolated" matters: with no bass or harmony distracting the ear, listeners hear deviation as error rather than feel. A humanize algorithm that pushes past the threshold stops sounding like a drummer and starts sounding like a malfunction.

Porter's onset-deviation study fills in where real lo-fi production actually sits. Across lo-fi breaks, the mean hi-hat onset SD and the mean velocity SD both land inside the human-timing band, and the timing figure stays well under the groove-perception ceiling. The lo-fi idiom, for all its reputation as loose or lazy, occupies the center of human timing tolerance rather than its sloppy edge.

The evidence converges. Repp's band says human deviation runs in a bounded range at typical hip-hop tempos; Senn's rating drop says listeners stop accepting randomization beyond that range; Porter's lo-fi corpus says released records actually sit near the center of the window. The variance-matching target in the decision rule is not a style preference — it is the parameterization of published timing research plus a corpus of the genre it is meant to serve.

The myth to drop is that swing equals human feel. Swing is a musical idiom — a long-short grid, not a randomness profile. The studies above describe distributions, clouds of deviations clustered around a mean. A swing preset does the opposite: it repeats the identical offset on every pass, which is why a "humanized" swing loop still locks into a mechanical clock pattern that listeners read as an artifact of the sequencer. Human timing is a variance problem; the fixed preset is a delay-line problem. That is exactly why the canonical rule permits the middle swing preset only for a deliberately stable boombap groove, and never as a humanize tool.

The Decision Table

Sort the standard production tempos by which humanize method survives a null test against a real drummer, and the pattern is unambiguous: AI randomize wins in most cases; the middle swing preset wins only when the reference loop's swing ratio is stable across every bar. The rows are not genre-mood categories; they are descriptions of what the target groove actually does over time.

The lo-fi hip-hop row is the row producers misread most. Lo-fi is not swung because a drummer plays a steady long-short pattern; it is swung because the hi-hat pull drifts from bar to bar. AI randomize under a hard timing cap with a modest velocity spread reproduces that drift, while the light preset's offset is too small to read as human and the heavy preset lands late before the first bar resolves.

In half-time trap, the groove lives in loose hi-hat clusters around the grid — notes slightly ahead, notes slightly behind, ghosts shaped by velocity. That is variance-matching work. A fixed swing preset shifts every long-short pair by the same interval, collapsing the cluster into a repeating limp. In drum'n'bass, tempo compresses inter-onset intervals, so a fixed swing offset that reads as style at slower tempi turns into mechanical phase accumulation; AI randomize keeps each note inside the perceptual window while letting the shuffle breathe.

In garage, the win comes from velocity. Ghost-note patterns sit in a low velocity band; the light preset's tiny offset does not separate them from the grid, and the middle preset locks the offbeat into a position garage rarely holds. The pop-rock row is the cleanest myth break: backbeats are straight, so the human element is snare velocity and a slight kick push, not a long-short ratio — swing presets have nothing to match.

The straight boombap row is the exception, and the constraint is the point. The middle swing preset is deterministic — it repeats the identical offset every bar — so it wins only when the reference loop's swing ratio is stable across every bar. That is the condition under which a deterministic delay line equals a human performance. The moment the loop drifts beyond that window, the preset's exact repetition reads as a clock artifact.

That leaves the light and heavy presets with no wins. The light preset's small offset falls inside the perceptual threshold where the ear still groups the note as grid-aligned — sloppy, not played. The heavy preset fails the opposite way: it crosses the groove ceiling, pushing the offbeat outside the accepted human-timing window. The working rule: AI randomize is the default; the middle swing preset is a conditional tool for stable boombap; the light and heavy presets are not humanize tools at all.

What the Data Doesn't Tell You

The timing cap is a compromise value, not a law of human timing. The loops behind the headline test were all recorded with a drummer alone in a room playing to a click, and that recording condition hides failure modes that the canonical rule does not anticipate.

First, a single global cap ignores instrument-level reality. In most onset-deviation datasets, kick and snare onsets cluster tightly, while hi-hats and rides scatter more widely. A single cap does different violence to each: it almost never binds on kick and snare, so the randomizer injects more deviation than a real drummer produces on those instruments — sloppy; and it truncates the long tail of hi-hat and ride timing — stiff. The rule's single cap is applied to what is really multiple distributions.

Second, the loops are solo performances. In a band, the hi-hat does not behave as an independent noise source; it moves as a correlated anticipatory gesture, pulling slightly forward before a chord change or a snare backbeat. Independent per-note sampling — the mechanism underlying AI randomize — cannot reproduce that correlated pull because each note is drawn without reference to its neighbors. You get the right average looseness with the wrong covariance structure.

Third, the tests that set the timing threshold ran on isolated loops, but isolation is the context where timing errors are maximally audible. In a full mix, masking from bass and vinyl noise widens the acceptable window considerably, so a deviation that breaks an isolated loop in a perceptual test can pass unnoticed in a real production. This cuts in favor of the rule: in dense lo-fi mixes, the cap errs toward conservative over-quantization rather than audible sloppiness.

Fourth, AI randomize treats notes as independent events, but human fills are gestures. A hi-hat run tends to accelerate toward the downbeat, and the acceleration is a sequence-level property — no per-note sample, VAE-based or otherwise, will generate it unless the model is conditioned on the whole run. There is a useful distinction from Research Randomizer's documentation: sorting generated numbers is helpful for random sampling but not desirable for random assignment. Per-note humanization is random assignment; an accelerating fill is a sorted structure, and no amount of independent jitter reconstructs it.

None of this is rescued by the other tools on the table, which is where the "swing equals human feel" myth collapses. Swing is a musical idiom, not randomness. A live drummer's swing ratio drifts from bar to bar, while a swing preset repeats the identical offset forever — the same long-short pattern at every grid position, which listeners read as a clock artifact. And a per-note VAE sample, like any independent draw, cannot create the sequence-level acceleration above. The swing preset that the decision table retains for a deliberately stable boombap groove is a musical idiom, not a humanize tool.

None of this overturns the decision rule; it maps its boundary. Use the rule's default for isolated loops and dense mixes, but hand-adjust when the kit has extreme SD asymmetry, when the groove was captured from an ensemble, when the arrangement is sparse enough that masking stops hiding errors, and when the part is a fill rather than a beat. In those cases the data stops telling you what to do — and that is exactly when you should trust your ears over the preset.

Worked Case

For a lo-fi or hip-hop session, the Stanford Microtiming Corpus track dusty_cans is the cleanest worked case for treating humanize as a variance-matching problem. Its hi-hat performance carries measured onset offsets that drift around the grid, with a mean near the grid and a modest SD. Measured against those offsets, the middle swing preset produces a large mean absolute error, while a capped Gaussian sampler lands close.

Applying the middle swing preset to the even positions puts every even-position note exactly the fixed offset late, no matter where the drummer actually played. The result is a constant long-short clock: the same delay repeats on every even subdivision, and that repetition is exactly the mechanical artifact listeners read as fake. The myth that swing equals human feel dies on this track, because the human offsets drift around the grid while the preset repeats a single fixed displacement.

Fitting a Gaussian to the measured offsets and then drawing multiple samples with AI randomize produces a small mean absolute error across all steps, with the best sample smaller still. The capped Gaussian works because it reproduces the shape of the human variance: some notes land early, some late, matching the corpus instead of displacing every even note by an identical amount.

Velocity is the other half of the variance-matching problem. The source hi-hat velocities form a varied human velocity profile. AI randomize with a modest velocity SD reproduces this distribution: a Kolmogorov–Smirnov test against the source returns a high p-value, so the sampled velocities are statistically indistinguishable from the human take. The middle swing preset leaves all velocities identical, which inverts the actual dynamics of dusty_cans, where consecutive hits differ in velocity.

The takeaway for a current mix session: default to AI randomize with a hard timing cap and a modest velocity SD for lo-fi and hip-hop humanization, and reach for the middle swing preset only when the target groove is a deliberately stable boombap pattern. dusty_cans is not that groove.

How to Choose Well

The useful question is not "how much swing does this need?" but "what variance does the reference actually carry?" The preset that sounds human is the one whose inter-onset-interval (IOI) distribution matches the recorded performance. That single framing resolves the endless DAW humanize debates — Digital Performer, Logic Pro, Reaper, FL Studio — at once. The rules below are a decision tree, not a menu.

Rule 1. If your reference drum performance was recorded by a human playing to a click, choose AI Randomize with a hard timing cap and a modest velocity SD. This is the default for lo-fi and hip-hop. The mechanism is variance matching: a click-constrained drummer produces an IOI distribution, and you want the generator to sample from that same distribution — not to impose a repeating offset. The default is DAW-agnostic in principle but not in practice. According to an Image-Line FL Studio forum thread (thread 244937, "How do you randomize/humanize midi note position timing"), FL Studio lacks a simple timing randomizer, so users there must fake the cap via the quantizer's random presets; Reaper users, by contrast, can build the equivalent by extending the existing MIDI velocity humanize function to item gain and note timing, as noted in Reaper forum workflow threads.

Rule 2. If the reference track's swing ratio is stable across all bars, use the middle swing preset alone. Do not add AI randomize on top unless the measured SD of the human reference is very low. The logic is strict: a swing preset repeats the same long-short offset on every iteration, and when the reference drummer was genuinely machine-stable, that repetition is exactly the target. Adding randomize on top smears the stability you just measured. Only when the reference SD is very low is the human playing like a sequencer anyway, so keep the preset clean and skip the randomize layer.

Rule 3. Never use the light or heavy swing presets as humanize tools. They are deterministic clock shifts, not randomness — treat them as stylistic "push" (the heavy preset drags the offbeat later) or "drag" effects. This is the myth kill: swing does not equal human feel. A live drummer's swing ratio drifts from bar to bar; a preset repeats an identical offset forever, and listeners read that mechanical long-short repetition as a clock artifact, not a performance. No amount of velocity variation rescues a deterministic shift.

Rule 4. If the track includes an acoustic ride cymbal or brush kit, skip both categories and record the part. AI randomize and the swing presets are trained on drum-machine and break contexts, not on idiomatic jazz ride patterns. An acoustic ride's timing is neither a variance problem nor a swing-ratio problem; it belongs to a different phrasal idiom that no randomize function will reproduce.

Rule 5. Before exporting, measure the treated MIDI's IOI SD across a passage. If it falls very low, raise the AI randomize cap; if it runs high, lower the cap and re-render. This is the verification gate that separates a chosen setting from a matched distribution. According to VI-Control forum discussions, the Humanize elements in Digital Performer can take a track from "sterile MIDI precision" to "dirty, flawed humanity" within 60 seconds — but sixty seconds of transformation still needs a measurement gate before it leaves the session.

Set the gate before you touch a preset: measure the reference's swing stability and IOI spread first, then apply the matching rule. The preset that survives the null test will be the one whose variance profile matches — never the one whose name sounds like "feel."

What to do next

Frequently Asked Questions

How many random-number sets has Research Randomizer generated since 2007, and what does its example output demonstrate?

Research Randomizer has generated more than 43.7 billion sets of random numbers since 2007, and its example output for five numbers in range 1-50 is 2, 17, 23, 42, 50 — no draw predicts the next.

What lower bound for human timing did Repp's meta-analysis establish, and why is it a generous floor?

Repp's meta-analysis in Psychonomic Bulletin & Review found trained musicians' tap intervals carry a standard deviation that scales with the inter-onset interval, and because those are finger-tap data rather than full-drum-kit performances, a real drummer should land at or above that band, making it a deliberately generous floor.

What happens to groove ratings when random microtiming deviations get large in isolated drum loops?

Senn et al. in Music Perception found that in isolated drum loops, large random microtiming deviations received lower groove ratings, so a humanize algorithm that pushes past that threshold stops sounding like a drummer and starts sounding like a malfunction.

Where do lo-fi hi-hat onset and velocity SDs sit relative to the human-timing band and the groove-perception ceiling?

Porter's onset-deviation study found that across lo-fi breaks, the mean hi-hat onset SD and mean velocity SD both land inside the human-timing band, and the timing figure stays well under the groove-perception ceiling.

Under what specific condition is a fixed swing preset acceptable as a humanize tool?

A fixed swing preset is only acceptable when the target is a deliberately stable boombap loop whose swing ratio barely moves, because the human target itself has almost no variance.

Why does a swing preset produce a clock artifact while an AI randomize engine does not?

A swing preset shifts an event, not a distribution, so the same offbeats move every bar and the long-short pattern loops as a fixed cycle, whereas the AI engine's output has no bar-to-bar correlation, so no such artifact can emerge.

Quick answers

Sources: arXiv, arXiv, Reddit, Reddit, arXiv