Tiny Timing Differences in Swing: Turning “Loose” into Audible Variables
Two jazz studies as a starting point for a proposed, untested synthetic listening exercise.
2026.11.07 · 7 min read · AI-assisted writing and translation
Swing is an experience. To investigate it, first make the timing changes explicit, then let listening answer the question.
Separate the meanings of “inexact”
When listening to jazz, phrases such as “loosen up” and “sit further back” are evocative, but they do not translate directly into production instructions. Does moving every note by a few random milliseconds add a human feel? A more useful question is: which note moves, relative to what, and in which direction?
There are at least four different operations. First, change the long-to-short ratio between successive eighth-note onsets within a beat: the swing ratio. Here, length means the time between onsets, not how long a key is held. Second, shift the entire solo part later, delaying both onbeat and offbeat notes. Third, add random microtiming deviations around each note’s intended position. Fourth, delay only the soloist’s onbeat notes relative to the accompaniment, while keeping the offbeats aligned. In this article, “onbeat” means the beat positions throughout a bar, not just its first beat.
All these operations can move waveforms away from a grid, but they change different relationships. Shifting an entire piece later would not change its internal rhythm at all. A delay needs a reference before it has musical meaning.
Synthetic onset diagram at 120 BPM. No audio or listening results are included. Delaying onbeats by 20 ms while retaining offbeats near 333 ms also changes the solo’s internal long-to-short intervals.
The two studies asked different questions
The 2019 study defined a grid using the average swing ratio, then manipulated small fluctuations around that grid. Removing those deviations did not reduce swing ratings; expanding them lowered ratings. This does not support the intuition that more deviation produces more swing, but it does not make every timing relationship irrelevant.
The 2022 study examined another structure: slightly delayed soloist downbeats combined with offbeats synchronized to the rhythm section. This arrangement enhanced swing in its listening experiment. The crucial difference is the manipulated variable. Fluctuations around average positions and systematic offsets between parts cannot usefully be lumped together as “inexact playing.”
Presenting the later paper as simply overturning the earlier one would miss the most useful lesson: a finding must be read alongside its definition of the variables. Neither result provides a universal millisecond prescription for every tempo, instrument, and style.
A listening exercise you can build
The following is a proposed teaching demonstration. It has not been run, and there are no listener data. Pair an original eight-bar melody with fixed synthesized accompaniment at 120 beats per minute. Keep pitches, timbre, velocity, note durations, mixing, and accompaniment identical across versions. Change only the solo part’s onset times, then check loudness after rendering so that louder is not mistaken for more compelling.
Baseline: each beat lasts 500 milliseconds. Place solo and accompaniment onsets at 0 and approximately 333 milliseconds, producing a 2:1 long-to-short interval ratio.
Ratio version: move only the solo offbeats to 250 milliseconds to create even eighth notes and make the ratio change audible.
Whole-part delay: shift every solo note in the baseline 20 milliseconds later, leaving the accompaniment unchanged.
Random version: starting from the baseline, independently shift each solo note by a random amount between −20 and +20 milliseconds. Fix the random seed and save the timing list.
Onbeat-delay version: shift only the solo onbeat notes 20 milliseconds later. Leave the offbeats at approximately 333 milliseconds, aligned with the accompaniment.
Twenty milliseconds is merely a starting setting for the demonstration, not a research-established optimum. One coupling must also be made explicit: the fifth version changes the solo’s internal long-to-short intervals. It cannot be described as changing only inter-part delay while holding the solo swing ratio completely constant. Nor is the fourth version equivalent to natural fluctuations in a real performance.
During production, use a short, fixed note duration to prevent shifted onsets from creating new overlaps. Keep the same leading silence and ending margin across versions, and remove condition information from filenames. Check the rendered onset times before listening, so that software humanization presets, changes in reverb, or accidental velocity edits are not mistaken for timing effects.
Record what you hear before revealing the labels
Compare the baseline with one variant at a time, playing anonymous A/B pairs in randomized order with the parameters hidden. Record three separate judgments: whether you can distinguish them, which swings more, and which you prefer. Allow “no difference.” Include identical-version pairs as a check, repeat with the playback order reversed, and test several seeds for the random version so the conclusion does not depend on one accidental pattern. Reveal the labels only after listening.
Even a stable preference would describe an experience with these synthetic excerpts, not establish the complete mechanism of real jazz. Investigating a broader effect would require more melodies, tempos, and listeners, plus comparison methods specified in advance.
The value of this exercise is turning a vague request for something “looser” into timing relationships that can be checked repeatedly. The next time a passage swings, ask: did the long-to-short ratio change? Did the whole part move back? Or did some notes arrive slightly late while others met again?