Piano performance, measured
624 recordings of 19 pieces were turned into lists of notes — every note's start, end, loudness, and both pedals — and compared. This page is the measurements first, then what they show. No musical training is assumed.
Each measurement below comes with two clips of the same performance of the same bars, with one quantity changed and nothing else touched. That is the point of them: if the two sound different, the difference is that one quantity. They are played by a computer from the notes, which is why they sound a little flat — that is deliberate, so nothing else distracts.
Here is the starting point everything is changed from: the notes played dead straight, and then the same bars set to the middle of what real pianists do.
After each pair there are the two real recordings that sit furthest apart on that measurement. They are harder to compare, because between two real recordings everything differs at once — which is the whole reason for measuring one thing at a time.
Does the music speed up and slow down across a long phrase, the way a sentence rises and falls — or does it hold one steady pace?
The second clip breathes: it presses forward in the middle of the phrase and eases at the end. The first stays level.
Within the phrase, does the player stretch and squeeze individual beats, or keep them even?
This is smaller and more local than the last one — single beats leaning long or short, not the whole phrase shape.
Below the level of the beat, are the fast running notes spaced evenly, like a machine, or slightly unevenly?
Listen to the rippling left hand underneath, not the tune. In the second clip its notes are not quite evenly spaced.
When one hand plays a steady accompaniment and the other plays the tune, does the tune land exactly with it, or drift behind?
The accompaniment keeps the same time in both. Only the tune moves — in the second clip it hangs back behind the beat.
When both hands are written to strike at the same instant, do they actually arrive together? Often the left hand lands first, on purpose.
A gap of a few hundredths of a second. It does not sound like two events — it sounds like a different kind of warmth.
Inside a single chord played by one hand, does the top note — usually the tune — sound a fraction before the notes underneath it?
This one is at the edge of what anyone can hear, which is worth knowing before trusting a description that claims to hear it. A deliberately exaggerated version is included so you know what to listen for.
Is a chord struck as one block, or spread out like a strum? And if spread, does it start from the lowest note or the highest?
Four clips: narrow against wide, then upward against downward. The direction is a much clearer difference than the width.
The number that comes back out of a rolled chord is about twice the setting that went in, because the measurement spans the whole chord and so also picks up whatever the two hands are doing apart from each other. These four captions name the setting, not a measurement.
How far above the other notes of its chord is the tune played — how much louder is the note you are meant to follow?
The first clip has the tune at exactly the same loudness as everything under it. It is the thing no pianist does, and hearing it explains why.
How far apart are the quietest and loudest notes — the distance the player uses between soft and loud.
Same notes, same speed. Only the distance between soft and loud changes.
Does each note carry on until the next one starts, so the line is joined up — or is there a gap between them?
Neither clip uses the pedal, so the fingers are the only thing joining the notes. The first leaves audible gaps; the second overlaps them.
The right pedal lets strings keep ringing after the fingers leave the keys. Is it lifted and re-pressed at every new chord, keeping things clear, or held across several so they blur together?
Three clips: no pedal at all, pedal changed at every chord, and pedal held so the harmonies run into one another.
The pedal is almost never pressed at the same instant as the chord. It comes just after, which catches the new chord and drops the old one. How long after?
Every pianist measured here does this. Only the delay differs, and it changes how cleanly one harmony gives way to the next.
Twelve of about thirty. The others are listed here in plain words; each links to its own page, which has the same two-ended audio.
| Measurement | What it is |
|---|---|
| Basic tempo | How fast the piece goes overall. |
| Linger | How long the longest hesitations are, measured against the player's own pulse. |
| Section contrast | Whether a player presses on into a contrasting section or draws back. |
| Coda fade | How much the closing bars are broadened relative to the music just before them. |
| Repeat fidelity | When a passage comes back, how closely its timing repeats the first statement. |
| Bass anticipation | How often the bass is placed clearly ahead of the melody. |
| Rolled fraction | How often block chords are rolled at all. |
| Roll direction | When a chord is rolled, whether it starts from the bottom or the top. |
| Third lift | The same lift, but measured only among the notes of the right hand. |
| Hand balance | How much louder the right hand is than the left. |
| Top-loudest rate | How reliably the top note is the loudest note of its chord. |
| Climax span | How much louder the piece's high point is than its opening. |
| High-loud slope | Whether higher notes are played louder. |
| Contour smoothness | How smoothly the melody's loudness changes from note to note. |
| Melody roughness | The opposite view of the same thing: note-to-note jumps in melody loudness. |
| Detachment rate | How often the melody is detached rather than joined. |
| Accompaniment legato | The same overlap measure, applied to the accompaniment. |
| Pedal density | How much of the time the sustain pedal is down. |
| Blur | How many changes of harmony are allowed to overlap under one pedal. |
| Soft-pedal share | How much of the time the una corda is down. |
| Emergent voice | How often a player brings out a line that the other players of the same piece leave inside the texture. |
A fair objection: maybe these thirty measurements are really three or four things wearing different labels — perhaps everyone who spreads chords widely also plays loudly. That can be tested by asking how much of the difference between recordings survives if you keep only a few combined measurements. The best single combination captures 18 % of it; the best two together 35 %. The rest is spread thinly across everything else.
So there is no small hidden set of master dials. To describe what separates one pianist from another you need most of the list.
Two versions of the question, with very different answers.
The easy one. Given a few seconds of a recording the computer has heard other parts of, name the pianist out of 27 candidates. Guessing gets 4 %. Four bars of music gets about 39 % if it may use every measurement — but only 8–14 % if restricted to one family at a time. No single aspect of playing carries someone's identity.
The hard one, and the real one. Work out a pianist's habits from their other recordings, then show the computer a piece by a different composer that this pianist has never been heard in, and ask who it is. Across 221 such tests it names the right person 43 % of the time, against 7 % for guessing — six times better than chance. Allowed three guesses, the right pianist is on the list 66 % of the time.
That is the substantive claim: a pianist's habits are portable enough to survive a change of composer. It is not reliable identification, and it is nothing like what an experienced listener does — but it is a long way from guessing.
For every pianist appearing in more than one piece, does being high on a measurement in one piece predict being high on it in another? The correlation runs from 0 (nothing carries over) to 1 (it carries over completely).
| Measurement | Carries over | Reading |
|---|---|---|
| Melody lead | +0.49 | travels with the person |
| Dynamic range | +0.39 | travels with the person |
| Roll direction | +0.38 | travels with the person |
| Hand dislocation | +0.34 | travels with the person |
| Melody float | +0.33 | travels with the person |
| Pedal changes / bar | +0.32 | travels with the person |
| Hand balance | +0.26 | partly |
| Roll width | +0.25 | partly |
| Pedal lag | +0.24 | partly |
| Melody lift | +0.23 | partly |
| Beat sway | +0.18 | set by the piece |
| Emergent voice | +0.17 | set by the piece |
| Soft pedal | +0.14 | set by the piece |
| Finger legato | +0.14 | set by the piece |
| Section contrast | +0.07 | set by the piece |
| Grain | +0.07 | set by the piece |
The pattern rearranges the usual vocabulary. What belongs to the person is how the notes of a chord come apart and how the two hands are weighed against each other. What is mostly decided for the piece includes the one that gets praised most — how far the tune is lifted above its chord. How evenly someone spaces fast notes carries over not at all.
Two of the weakest entries here are measurements the transcription is bad at (see Part 4). A measurement made badly cannot show a habit even when there is one, so for those the honest conclusion is "this cannot tell you", not "it does not carry over".
For each piece: how far apart its recordings sit, in real units, against how far apart pianists sit in general. 1.0 would mean "as varied as pianists anywhere"; 0.5 means these players agree twice as closely.
| Most varied | Least varied | ||
|---|---|---|---|
| Bach Prelude in C | 0.64 | Schubert Op. 142/3 | 0.30 |
| Träumerei | 0.63 | Chopin Ballade 4 | 0.32 |
| Nocturne Op. 9/2 | 0.62 | Chopin Étude 10/1 | 0.35 |
The whole range is narrow — every piece leaves roughly the same amount of room. But the order runs against intuition: the pieces at the bottom are the hard ones, and the pieces at the top are ones a student meets in the first year or two. The clearest version is inside one composer — Ravel wrote both the Pavane, which a beginner can hold, and Ondine, which pianists work at for years. They sit two places apart out of nineteen.
One explanation is attention: a piece that does not threaten the notes leaves the player nothing to do but decide how it should go. Another is simpler — a hard piece is hard partly because the composer specified a great deal, and specification is what leaves no room. Nothing here separates the two.
A spread measured on few recordings reads low. The three at the bottom have the fewest recordings; corrected for that they come out level with each other rather than lowest. The direction of the finding survives, but no single position in the middle of the table should be read closely.
83 of the recordings are by people who are not professionals. The expectation is that they differ the way a beginner's handwriting differs from a calligrapher's: the same letters, worse. That is half right.
The half that holds is scatter. Their note-to-note consistency is far wider — lead scatter 2.4×, pedal changes / bar 2.3×, melody roughness 2.2×, articulation scatter 2.1× the professionals' spread.
The half that fails is the expressive measurements. On those the home recordings are not more varied but less, and sit toward the modest end of every one: narrower between soft and loud, less lift on the tune, less separation between the hands. They do not have a different vocabulary. They use less of the same one, with more noise around each word.
On this evidence skill is not a style. It is the tightness of the scatter around a style, and separately the confidence to move a measurement far from the middle.
111 of the recordings did not come from listening to audio at all. They are from a competition whose entrants play a piano with a sensor under every key and pedal, so the instrument wrote the file as it was played.
A transcription reports the sustain pedal as a switch — down or up. The sensors report its position continuously, and show these pianists holding it part-way down about 26 % of the time: catching some of the ringing and letting the rest go. It is something pianists discuss constantly and it is entirely absent from every audio transcription here, which has only two positions to describe it with.
So the pedal measurements on this page are real, but they are a continuous gesture seen through a switch.
One pianist is in the corpus twice: as a fifteen-year-old competitor in 2013 and in commercial recordings of the same two pieces in 2018. What did not move in either piece is the group of habits above — how far the tune is lifted, how the hands are weighed, how much the fingers overlap, how often the foot changes (identical to one decimal place, five years apart).
What moved, and moved the same way in both pieces: slower, chords spread wider, hands closer together, pedal later, phrases arched a little more. A second pianist, recorded fourteen years apart, shows the same split — but moved in the opposite direction on two of them. Whatever the years did, it was not a general law of growing up.
Two pianists prove nothing general. It is the shape a real answer would have.
The measurements come from software listening to a recording. That can be checked: a sensor recording was played back through a synthesiser, handed to the same software as if it were audio, and re-measured — the same performance, once known exactly and once guessed. The error is shown below as a fraction of how far pianists actually differ. Below about 0.4 the error is small next to the thing being measured; above 1.0 the error is bigger than the signal.
| Measurement | Error ÷ real spread | |
|---|---|---|
| Bass anticipation | 0.04 | trustworthy |
| Basic tempo | 0.06 | trustworthy |
| Linger | 0.06 | trustworthy |
| Arch depth | 0.06 | trustworthy |
| Section contrast | 0.07 | trustworthy |
| Melody float | 0.09 | trustworthy |
| Coda fade | 0.13 | trustworthy |
| Beat sway | 0.16 | trustworthy |
| Repeat fidelity | 0.27 | trustworthy |
| Melody lead | 0.32 | trustworthy |
| Roll direction | 0.35 | trustworthy |
| Hand dislocation | 0.35 | trustworthy |
| Roll width | 0.39 | trustworthy |
| Emergent voice | 0.51 | usable |
| Pedal density | 0.54 | usable |
| Rolled fraction | 0.57 | usable |
| Pedal lag | 0.73 | usable |
| Third lift | 0.86 | usable |
| Grain | 1.00 | do not rely on |
| Melody roughness | 1.07 | do not rely on |
| Soft-pedal share | 1.20 | do not rely on |
| Blur | 1.24 | do not rely on |
| Top-loudest rate | 1.31 | do not rely on |
| Pedal rate | 1.43 | do not rely on |
| Accompaniment legato | 1.60 | do not rely on |
| Finger legato | 1.60 | do not rely on |
| Melody lift | 1.62 | do not rely on |
| High-loud slope | 1.62 | do not rely on |
| Climax span | 1.95 | do not rely on |
| Detachment rate | 1.96 | do not rely on |
| Hand balance | 1.98 | do not rely on |
| Dynamic range | 2.08 | do not rely on |
| Contour smoothness | 2.47 | do not rely on |
Everything built from when a note starts survives almost intact — the timing measurements on this page are essentially exact. Everything built from when a note stops, how loud it was, or where the foot was is worse than the differences it is being used to measure.
Two things soften that and one sharpens it. The errors are mostly consistent biases rather than random noise — the software compresses everyone's loudness by about the same amount — and a bias shared by everyone largely cancels when the question is who is wider than whom. But noise also pushes correlations toward zero, so where a badly-measured quantity still showed a habit carrying over, the real effect is probably stronger than stated.
This test is a floor, not an estimate. The audio was synthesised cleanly, with no room, no audience and no tape hiss, so real recordings are harder than this and the true errors are larger.