Rubato, voicing, touch: the pianist's vocabulary, measured

If you are looking for the chess variant previously hosted here, it has moved to siegechess.com.

Here are two pianists, recorded sixty-eight years apart, playing the first two bars of Schumann's Träumerei. They take exactly the same time over it — 9.2 seconds each — and play at the same loudness.

They do not sound alike, and none of the difference is speed or volume. One keeps the right pedal down for almost the whole passage; the other lifts it for more than a third of it. One bends single beats more than twice as far as the other. One lets the tune sit further above the chord under it, and strikes it further ahead of the other notes. And here is the same passage with none of that — the notes, in time, at one volume:

To describe the difference between the first two, a pianist has about five words. Rubato: bending the speed, taking time from one note to give to another. Voicing: making one note of a chord — usually the tune — louder than the notes under it. Touch: how the key is pressed, and the tone that is supposed to result. Articulation: how long each note is held before the finger lets go. Pedalling: when the right pedal lifts the dampers off the strings, and how much of one chord is allowed to bleed into the next.

There is nothing wrong with those five words. Each points at something real, and teachers have used them for three centuries because they work. What they have never had is a number. They stayed vague because the measurement was hard: you need the exact start, end and loudness of every note, plus the pedals. Until recently that meant a concert grand full of sensors, or a graduate student with a waveform editor — Bruno Repp measured 28 recordings of one Schumann piece that way in 1992, at two hours each.

Machine learning removed that obstacle. Every recording here went through Transkun (Yan & Duan, ISMIR 2024), which turns a piano recording into a list of notes with loudness and both pedals, and then through parangonar, which ties every played note to the note the composer wrote. Once the two lists are tied together, everything below is arithmetic.

What is in the corpus

624 recordings of 19 pieces, from Bach to Aphex Twin, chosen for the variety of what the hands have to do — bells, chorales, a tune over a running left hand, pieces that change gear in the middle, one that never leaves 79 beats a minute, one that is nothing but broken chords.

They come from three places, and it matters which. Most are released recordings — professional ones, uploaded to YouTube by their labels; they are the "Professional" column in the table below. 83 are home recordings — videos by people who are not professionals; they are labelled wherever they appear and get their own section. And 111 are of a third kind: sensor recordings from the Minnesota International e-Piano Junior Competition, whose entrants play a Yamaha Disklavier, a concert grand with a sensor under every key and every pedal. Those files were not transcribed from audio at all. The piano wrote them down as it was played, which makes them the only recordings here whose numbers are measured rather than inferred — and, later on, the only way to check how far the inferred ones can be trusted.

Throughout, "the recordings of a piece" means its released and sensor recordings together; home recordings are always counted separately and never enter a range, an extreme or a consensus.

The consensus performance is sometimes rather good and sometimes not, and which one you get depends on the piece. Where the pianists broadly agree about where a phrase should rise and fall — as they do in the Träumerei below — the middle of them is a perfectly musical reading, and there is no reason it should not be. Where they disagree, the middle commits to nobody's decision and can sound like a very careful sight-read. What matters here is not that it is good but that it is neutral: every dial sits at the median of that piece's recordings, so when one number is changed it is the only thing there is to hear. Here it is for the passage above:

The dials, one at a time

The five words turn out to name about thirty measurable habits, or dials. The twelve most important are walked through below, one piece each, so that the whole corpus gets heard. Each dial gets a plain description, a precise definition in a box, a grey pair with only that dial changed, and then the two real recordings of that piece that sit at its extremes. Overall speed is deliberately not one of the dials: that one pianist is slower than another is the least interesting thing about them.

Rubato is four numbers

Rubato is not one thing. A pianist can bend the speed across a whole phrase, stretch single beats inside a bar, space the notes unevenly inside a beat, or let the tune drift against a steady accompaniment — and each of these is independent of the others. That they are separable is not a modern discovery. In 1913 the young pianist Maurice Dumesnil took Clair de lune to Debussy for a lesson and was told to stop pulling individual beats about: rubato, the composer said, "must be done within the entire phrase, never on a single beat" (Dumesnil wrote the lesson down years later). Debussy was drawing a line between the first two of the four numbers below — approving of one, forbidding the other.

1 · Arch depth — Schubert, Impromptu in G-flat

In plain words: does the speed rise and fall across a phrase, like breathing, or stay level? Schubert's Impromptu is one long tune over sextuplets that never stop, and its 43 recordings arch it by anything from 10 % (Eric Lu) to 28 % (Maria João Pires). Both grey clips are the consensus performance; only the arch depth differs.

And the two real recordings at those extremes. Harder to hear, because now everything else differs too — that is the whole problem with listening for one quality at a time:

2 · Beat sway — Beethoven, Sonata "Pathétique", slow movement

In plain words: inside the phrase, does the pianist stretch or squeeze single beats? This is the one Debussy told Dumesnil to leave alone: arch the phrase, he said, but do not lean on the beats inside it. Beethoven marks this movement Adagio cantabile — slow, singing — and its 47 recordings sway its beats by anything from 5.3 % (François−René Duchâble) to 11.3 % (Rudolf Buchbinder).

3 · Grain — Mozart, Sonata in C, K. 545

In plain words: below the beat, are the fast notes evenly spaced? The left hand of Mozart's "easy" sonata is an Alberti bass — the same four sixteenth notes over and over — and every piano teacher listens for exactly this in it. Its 37 recordings run from 4.3 % of the beat (Hans Leygraf) to 10.7 % of the beat (Carl Seemann).

4 · Melody float — Chopin, Fantaisie-Impromptu

In plain words: the older meaning of rubato — the tune bends while the accompaniment does not, "the left hand is the conductor", as Chopin is supposed to have said. It needs a steady accompaniment to measure against. The slow middle of the Fantaisie-Impromptu is exactly that: a singing tune over a left hand that keeps going. Its pianists float the tune both ways, from −44 ms (Vladimir Feltsman) to 90 ms (Valentina Lisitsa) — a negative number means the tune arrives before the point the left hand implies, a positive one after it.

"Together" is a spectrum

Notes written as simultaneous are almost never played simultaneously. Caroline Palmer measured this in 1989 and Werner Goebl pinned it down in 2001 with 22 pianists on a sensor-equipped Bösendorfer: the tune note of a chord sounds about 20–30 ms before the others. Most of that lead is physics — a key struck harder reaches the string sooner, so whoever plays the tune louder gets the lead for free. Separately, there is an old and deliberate habit — Neal Peres da Costa calls it dislocation — of putting the bass down before the tune, heard on almost every recording made before 1930 and long thought extinct.

5 · Hand dislocation — Satie, Gymnopédie No. 1

In plain words: does the left hand land before the right, or after it? Dislocation is largest where the texture is simplest, and Satie's is the simplest there is: a bass note, a chord, a tune. Its 23 recordings run from −49 ms (Bertrand Chamayou) to 58 ms (Roberto Prosseda) (negative = left hand first).

6 · Melody lead — Chopin, Nocturne in C minor

In plain words: within one hand, does the top note of the chord come a hair before the rest? The opening of the C-minor Nocturne is a chorale, every beat a chord with the tune on top. Its 27 recordings lead the tune by 0.0 ms (Zheyu Li) to 9.4 ms (Sheng Cai) — and both are near the edge of what anyone can hear, which is worth testing on yourself before trusting anyone who claims to hear it:

7 · Roll width and direction — Rachmaninoff, Prelude in C-sharp minor

In plain words: is a chord struck as a block or spread like a strum — and if spread, from the bottom up or the top down? The bells that open Rachmaninoff's Prelude are three-note chords over octaves, and the roll is the first thing you hear about a pianist there. Its 29 recordings spread their widest chords by 34 ms (Dubravka Tomšič) to 76 ms (Boris Giltburg), and start them from the bottom 3 % (Boris Giltburg) to 38 % (Marta Deyanova) of the time.

The four clips above are the one place where the caption gives a dial setting rather than a measurement. Roll width is defined as the first-to-last onset of a whole chord, which also contains whatever the two hands are doing apart from each other, so the number that comes back out of a rolled chord is about twice the number that went in. Every other grey caption on this page states a value that measures back off the clip: melody lead, hand dislocation and melody lift to the decimal place, and dynamic range to within a few velocity units — the last one only approximately, because stretching a phrase's dynamic arc cannot land a percentile spread exactly where you ask.

Voicing, in velocity units

"Voice the tune" means make it louder than the notes around it, and a transcription gives that directly. Loudness comes as MIDI velocity, a number from 0 to 127 for how hard the key was struck. It is roughly logarithmic — about a quarter of a decibel per unit — and it depends on how the recording was mastered, so it is only comparable within one recording. Everything here is therefore a difference in velocity units.

8 · Melody lift — Brahms, Intermezzo in A

In plain words: how far above the chord does the tune sit? Brahms's Intermezzo is the piece pianists reach for to show what an inner voice is, and its 28 recordings lift the top by 13 (Michael Shilyaev) to 26 (Wilhelm Kempff) velocity units. Zero is included below because it is the thing nobody does, and hearing it explains why.

9 · Dynamic range — Debussy, Clair de lune

In plain words: how far apart are the softest and loudest notes? The climax of Clair de lune is where it shows most. Its 27 recordings span 39 (Monique Haas) to 63 (Lang Lang) velocity units.

Touch, and the pedal

Touch is the word that has to stay mostly unopened. Everything the finger does before the hammer flies — pressed versus struck, the shape of the key's descent, the finger-key noise Goebl showed is what listeners actually use to tell a struck tone — is invisible in note-level data. What survives is how long the note is held.

These two dials are the ones where the transcriptions are weakest, and it is worth saying so before the numbers. A model listening to a recording can hear a note start; it cannot really hear a damper come off a string, and it cannot hear a pedal that is held halfway down. The 111 sensor recordings described above are the check on that, because on those the pedal was not estimated — it was measured, 11.3 times a second, as a number from 0 to 127. What they show is that the transcriptions get the pattern of pedalling roughly right and the detail not at all: 26 % of the time a sensor recording has its sustain pedal somewhere between fully up and fully down, a state the transcriptions almost never report. So read pedal changes per bar and pedal lag as real but coarse, read finger legato as relative, and treat everything about half-pedalling as coming from the sensor recordings alone.

10 · Finger legato — Bach, Prelude in C

In plain words: do the fingers connect the notes, or leave a gap between them? Bach's Prelude in C is nothing but broken chords in sixteenth notes, and how connected to play them is the oldest argument in Bach playing — some pianists here use no pedal at all, so the fingers are all there is. Its 27 recordings run from −0.02 (Glenn Gould) to 1.30 (Pierre−Laurent Aimard) (a negative number is a gap, as a share of the space between the notes; positive means the keys overlap). The two grey clips have no pedal:

And the two real recordings at those extremes:

11 · Pedal changes — Chopin, Nocturne in E-flat

In plain words: is the pedal changed with every new chord, or held across several so they blur together? The left hand of the E-flat Nocturne plays a bass note and two chords per beat, and its 32 recordings change the pedal anywhere from 2.5 per bar (Daniel Barenboim) to 4.5 per bar (Aldo Ciccolini).

12 · Pedal lag — Aphex Twin, Avril 14th

In plain words: every pianist here presses the pedal a moment after the chord — the "syncopated" pedal every teacher drills, which catches the new chord and lets go of the old one — but the moment varies. Avril 14th's recordings, the composer's own piano-roll take among them, lag by 98 ms (Josh Cohen) to 204 ms (Olga Scheps).

What the reviewers heard

Here are real sentences from published reviews of five of the Debussy pianists, next to what the measurements say about the same thing.

The reviewers are not wrong, and neither is their vocabulary — every one of those sentences is about something the numbers can find. What the words cannot do on their own is say how much, or which kind. "He doesn't add unnecessary rubato" is true of Gieseking, but by itself it cannot tell you whether he is steady across the phrase, steady inside the beat, or steady between the notes of a beat. He is all three, and the amounts are in the right-hand column.

A pianist as a column of numbers

Put it together and each recording becomes a column of numbers. To show what that looks like, here is one piece done exhaustively: every released recording of Clair de lune in the corpus, 26 of them, oldest on the left and newest on the right. There is nothing special about this piece — the same table exists for the other twelve, and they are all in the appendix. It is simply the one with a long enough run of recordings, 1952–2021, to be worth reading left to right.

Are these 27 dials really 27 different things? It is a fair question. Some of them could be one habit measured twice: if everybody who spreads their chords wide also plays loudly, then "spread" and "loud" are a single dial wearing two labels, and the real list is shorter than it looks.

That can be tested. Ask how much of the difference between recordings you can still capture if you are allowed to keep only a few combined measurements instead of all 27. The best single combination accounts for 18 % of it, and the best two together for 35 % — about a third. The other two-thirds is spread thinly across everything else. To capture most of what separates one pianist from another, you need most of the 27.

So there is no hidden pair of master dials of which rubato, voicing, touch, articulation and pedalling are elaborate restatements. The pianist's vocabulary was not too coarse after all. It was about the right size, and had no units.

Can the numbers tell who is playing?

That question has an easy version and a hard one, and they give very different answers. The easy version is the one a computer is good at: here is a passage from a recording the machine has already heard other parts of — whose is it? The hard version is the one a listener actually faces: here is a pianist in a piece you have never heard them play — do you recognise them? Both are below, in that order.

The easy version: the same piece

All 27 Clair de lune recordings were chopped into short stretches, the dials measured on each, and a classifier asked to name the pianist from a stretch it had never been shown. Every one of the 27 pianists is a possible answer, so guessing gets 4 %. How well it does depends almost entirely on how much music it is given.

Four bars is close to the hardest form of the question, and it is the one worth noticing: even a few seconds is enough to beat guessing many times over. Give it most of a page and it is right about half the time out of 27 possible answers. Nobody's identity lives in one family of dials either. On the 4-bar test, articulation on its own gets 8 %, timing on its own 10 %, loudness on its own 14 % — every one of them far above guessing, every one far below the 32 % that comes of using all of them at once.

The hard version: a piece they have not been heard in

That test had it easy: the machine had already heard the same pianist play the same notes a few bars earlier. Nothing about it says the numbers describe a person rather than one afternoon in a studio. The real question is whether a habit shown in one piece is still there in another.

48 of the pianists here recorded at least three of the pieces, which makes a clean test of exactly that. Hide one piece. From the pianist's other recordings, work out their signature: for every dial, how far they usually sit from the other pianists playing the same music. Now take a recording of the hidden piece — music this pianist has not been heard in — and ask which signature it sits closest to. The answer has to be chosen from everyone who recorded that piece: 6 candidates for the thinnest piece, 22 for the best-covered, 18 in the middle. Repeat for every pianist and every piece in turn, and that is 221 recordings judged.

It names the right pianist 43 % of the time. Guessing would get 7 % — so it is about 6 times better than chance, from nothing but the numbers, in music it had never heard that person play. Widen it to a shortlist of three and the right pianist is on the list 66 % of the time. It is not identification you would want to rely on, and it is nothing like what a human expert does. But it is not close to guessing either, and that is the whole claim: a pianist's habits are portable enough to survive a change of composer.

Which habits? Below, each dial is used on its own — the whole judgement resting on one number — against what guessing would give.

Not one of them is close to sufficient alone, which is why the full test uses all of them together. But the order is the interesting part, and it agrees with the section that follows: the best single habit is melody lead (17 %), and the ones at the bottom, ending with linger (7 %, at or below chance), carry almost nothing about who is playing. Some habits really do travel with the person from Bach to Rachmaninoff; others are decided fresh for each piece.

If you would like to hear what a signature sounds like rather than read it, the appendix has three pianists playing the same passage of each of six pieces.

What travels with the pianist

96 pianists appear in more than one piece here, 48 of them in three or more. That makes a general version of the test above possible: express every recording as a z-score among the other recordings of its own piece — how far this pianist sits from everyone else playing the same notes — and ask, over every pair of pieces a pianist appears in, whether being high on a dial in one piece predicts being high on it in the other. The number is a rank correlation, ρ, from −1 to 1: near 1 means the dial travels with the person, near 0 means it is set by the music, the instrument or the day. p is the chance of seeing a correlation that big by luck alone. With a few dozen pianists this is an indication rather than a measurement.

The result rearranges the five words. What travels with the pianist is, first, everything about how the notes of a chord come apart and how the hands weigh against each other: melody lead (ρ = +0.49), roll direction (+0.38), hand dislocation (+0.34), hand balance (+0.26), melody float (+0.33) and roll width (+0.25). A pianist who spreads chords wide in Brahms spreads them wide in Chopin. Dynamic range travels about as strongly (+0.39), and so, more weakly, do the pedal habits — how often the foot changes (+0.32) and how late it comes down (+0.24). What travels least is more surprising. Melody lift — the voicing number, the one reviewers praise pianists by name for — travels only weakly (+0.23); a pianist who lifts the tune high above the chord in one piece is only a little more likely than anyone else to do it in the next. Beat sway (+0.18) and finger legato (+0.14) are weaker still, and grain (+0.07) does not travel at all: how evenly a pianist spaces the fast notes in one piece says nothing about the next. One caution on that last group, which is settled further down: beat sway is measured almost perfectly, so its weakness is real, but finger legato and grain are two of the dials the transcriptions are worst at, and a dial measured badly cannot show a correlation even if one is there. On this evidence: the vertical is the person; the horizontal and the loud-soft of the tune are mostly the piece. How the hands and fingers arrive relative to each other, and how the two hands are weighed, is a habit of the hands, carried from work to work. Rubato and voicing are readings — chosen for the piece, and chosen again for the next one.

How much expression does a piece invite?

Do some pieces make pianists differ from each other more than others? Within one piece a z-score cannot answer that (it is scaled to the piece's own spread). What can is the spread in the dials' own units — milliseconds of lead, velocity units of lift — because those mean the same thing in any music. For each piece and each dial, take the spread among its recordings (the middle half of them, so one eccentric does not decide it) and divide by the same spread pooled over every professional in the corpus. 1.0 means "as varied as pianists in general"; 0.5 means these players agree twice as closely as pianists in general do. Home recordings and opening-only recordings are left out of this, as is basic tempo (its unit is the piece's own beat), arch depth (a Lento that turns into an Agitato is arch depth) and section contrast (whether the fast section is three times as fast is the composer's decision).

Two things come out of this, and the second is the interesting one.

First, the totals are close together: from 0.30 (Schubert Op. 142/3) to 0.64 (Bach Prelude in C), a little over a factor of two across 19 pieces and three centuries. Whatever a piece is, it seems to come with roughly the same amount of room in it.

Second, and this is the part that runs against the intuition: the order has nothing to do with difficulty, and if anything it runs backwards. The ranking, most varied first, is Bach Prelude in C and Träumerei at the top and Schubert Op. 142/3 and Chopin Ballade 4 at the bottom. The bottom three are all pieces written to be hard: Chopin's Étude Op. 10 No. 1 (0.35), his fourth Ballade (0.32) — which with Ondine is the hardest music here — and Schubert's B-flat Impromptu (0.30). Nor is the étude's placing a quirk of who plays it. Its 31 recordings are not only competition entrants; Arrau, Pollini, Ashkenazy, Biret, Perahia and Lisiecki are in there too, six pianists spread over four decades and four traditions, and the piece still leaves them as little room as anything here. Against it: Satie's Gymnopédie (0.60), Schumann's Träumerei (0.63) and Bach's Prelude in C (0.64), three pieces a student meets in the first year or two, at the top. Avril 14th (0.60), two minutes anyone could learn in an afternoon, sits well above Chopin's Fantaisie-Impromptu (0.42). The tidiest version of the test sits inside one composer. Ravel wrote both the Pavane (0.44) — a tune, a few chords, and a tempo a beginner can hold — and, nine years later, Ondine (0.42), which pianists spend years on before playing it in public. About as far apart in difficulty as two pieces by one man get; two places apart in nineteen. The pattern is a tendency and not a law, and the other étude is where it frays: the "Revolutionary" (0.47) lands mid-table, above Mozart's K. 545 (0.41) and the Pathétique's Adagio (0.42), both a great deal easier to play. One caution about the ends of this ranking, which is where its argument lives: a spread measured on few recordings reads low, because the quartile range of nine recordings recovers about 85 % of the true one against 96 % at thirty. So the Ballade (12 recordings) and the Impromptu (9) sit lower here than the music warrants, and Avril 14th, whose figure rests on 5 released recordings, sits higher. Corrected for that the three at the bottom come out level with one another, at about 0.36, and Avril climbs near the top — the direction of the finding survives, but no single place in the middle of the table should be read closely.

One reading is that this is about attention rather than about the music: a piece that is technically undemanding leaves the pianist nothing to do but decide how it should go, while a piece hard enough to threaten the notes gets played the way it is written, because that is the whole job that evening. Another is simpler — a hard piece is hard because the composer specified a great deal, and specification is exactly what leaves no room. Nothing here can separate those two, since the data has no measure of how hard a pianist found a piece; take it as a finding in search of an explanation. What the data does show is that "simple" and "interpretable" are not opposites, and may even point the same way.

Underneath the totals, pianists disagree about different things in each piece, and that is where the texture shows:

Read down a column and you get the piece's personality. Bach's Prelude has no chords to roll and no tune to lift, so all of its room is in the things it does have — pedal, articulation, grain, dynamics — and its pianists use it. Satie's Gymnopédie, the piece with the fewest notes, leaves the most room on everything that is not a note — hand balance, pedal lag, melody lift, beat sway — because when there is one bass note, one chord and one tune, the balance between them and the moment the foot moves are the performance. Avril 14th is the same lesson from the other side: its pianists agree almost perfectly on the things the piece nails down (roll width, hand dislocation, tempo shape, the composer's piano roll at one end and everyone else near it) and disagree wildly on hand balance, the one dial its texture leaves open. The Rachmaninoff leaves its room in grain (the fast middle section), the Brahms and the Debussy in how the hands come apart on the big chords, the Fantaisie-Impromptu in melody float. So the question is not how much room a piece leaves, but where — and "simple" pieces do not leave less. They leave it somewhere else.

Habit and decision

Every number so far is an average, and an average is a habit: how much louder the top usually is, how far the left hand usually lands ahead. There is a second kind of thing a pianist does that an average cannot see. Olga Scheps's Avril 14th is a clean example. She plays it faster than anyone else in the set — 93 beats a minute against a corpus that sits at 79 like a metronome, the composer's own take included — and her habits are otherwise unremarkable for the piece. What she does that nobody else does is, at a couple of places, decide that a different line is the tune: she brings the bass up under the second theme until it is the thing you hear, and near the end she leans on an inner note and lets the written melody step back. A bar of that moves a piece-wide average by a fraction of a unit. It is a decision, and decisions want counting, not averaging.

Counted that way, everyone has a few events, because the measure has a floor of ordinary unevenness. What separates Scheps is the size and the extent: her strongest departure is 28 velocity units above the consensus, the largest in the set, and events cover 56 % of her bars against 39 % for the composer's own take. Bars 8–14 are her biggest: the tune drops into the middle of the keyboard, everyone plays it louder there, and on top of that she brings the bass up by 20–28 units more than anyone else, so that for six bars the left hand's lowest notes are a second melody. Bars 9–11, the composer against Scheps:

Home recordings

83 of the recordings are by people who are not professionals, and the natural expectation is that they differ from the professionals the way a beginner's handwriting differs from a calligrapher's: the same letters, worse. Compared dial by dial inside each piece that has both, that expectation is half right. The table shows, for each dial, how far the home recordings' middle sits from everyone else's middle (in units of everyone else's spread) and how much wider their spread is (a ratio; 2 means twice as spread out). "Everyone else" here means the released and sensor recordings together — so the comparison group includes the competition entrants, who are not professionals either, but who can certainly play.

The half that is right is scatter. On almost every within-performance spread the home recordings are wider — the scatter of their melody lead several times the professionals', melody roughness (how much the tune's loudness jumps from note to note) three times, articulation scatter two and a half times — and they change the pedal a great deal less often, with some barely using it at all. That is the beginner's handwriting: the same marks, less control over each one. The half that is wrong is the expressive dials. On those the home recordings are not more varied but less: hand dislocation, melody float, soft pedal and pedal lag are all narrower, and their middles sit toward the modest end of every one — dynamic range narrower, melody lead shorter, lift lower. Home pianists do not have a different vocabulary; they use less of the same one, less deliberately, and with more noise around each word. Skill, on this evidence, is not a style. It is the width of the scatter around a style, and separately the confidence to move a dial far from the middle.

Three pairs, the same bars from a released recording and a home one. Listen for the scatter rather than the choices: the notes that are late by accident, the accompaniment that pokes out where nobody meant it to.

What a wired piano knows

Everything so far rests on a machine listening to a recording and writing down what it thinks it heard. That is worth checking, and 111 of the recordings — the ones from the e-Piano Junior Competition, spread over 14 of the pieces — make it possible, because they did not come from listening. Their pianos took notes.

One expectation to get rid of first. A sensor recording is not a tidier performance. Measured against the printed score the two kinds of file are much alike — a transcription links 96 % of the written notes, a sensor recording 95 % — because what is missing from a live competition round is mostly the pianist's doing: a different edition, an ornament taken differently, the odd note not played. What a capture gives is not a cleaner performance. It is precision about the notes that are there.

Start with what is simply in the two kinds of file. A transcription reports the sustain pedal as a switch: down or up, and nothing else. The Disklavier reports the position of the pedal as the foot moves it.

The last row is the one to look at. About 26 % of the time — a quarter of every performance — these pianists' right feet are somewhere in the middle — half-pedalling, catching part of the resonance and letting the rest go, riding the damper against the string. It is a thing pianists talk about constantly and it is completely absent from every transcription in this corpus, which has only two positions to describe it with. The same is true of the left pedal: the Disklavier records about 101 distinct soft-pedal positions where a transcription has 2. So the pedal dials in this post are real, but they are the shape of a continuous gesture seen through a switch.

Then the harder question: on the dials the transcriptions do report, how wrong are they? That can be measured, because a sensor recording can be played through the same sampled piano the clips above use, handed to the transcriber as audio, and the result compared with the file it came from. Same performance, once as ground truth and once as a guess.

The last column is the one that matters: the transcription's average error divided by how far that piece's pianists actually differ from each other on that dial. Below about 0.3 the error is a rounding error next to the thing being measured. At 1.0 the error is the size of the signal.

The pattern is sharp, and it is the one the rest of the post has been assuming. Everything built out of when a note starts survives almost intact — arch depth and beat sway come back at 0.06 and 0.16, which is to say the rubato numbers in this post are essentially exact. The chord-timing dials are next, at 0.27 to 0.31: good enough to rank pianists, not good enough to quote to the millisecond. Everything built out of when a note stops, how loud it was, or where the foot was is worse than the differences it is being used to measure — pedal changes per bar 1.41, dynamic range 2.04, finger legato 1.58, hand balance 1.95.

Two things soften that, and one thing sharpens it. Softening: the errors are largely biases rather than noise — the transcriber compresses every performance's dynamic range by about the same amount, and puts every pedal down a little late — and a bias that is the same for everyone mostly cancels when the question is which pianist is wider than which. And the dials with big errors are exactly the ones this post already treats gently. Sharpening: measurement noise pushes correlations toward zero, so where a noisy dial still showed a pianist's habit travelling from piece to piece — dynamic range, hand balance — the real effect is probably stronger than the number given above; and where a noisy dial showed nothing, like finger legato, the honest conclusion is not "it does not travel" but "this cannot tell you".

Two things this test is not. It is a floor, not an estimate: the audio was synthesised from a clean sample set with no hall, no audience and no tape hiss, so a real recording is harder than this and the true errors are larger. And it says nothing about the sensor recordings themselves being right — a Disklavier is a measuring instrument, but it is measuring its own action, not a Steinway in Vienna.

The same pianist, years apart

One thing these recordings gave that nobody planned. Eric Lu is in the corpus twice over: as a fifteen-year-old competitor in 2013, playing Schubert's Impromptu and the Brahms Intermezzo on the competition's Disklavier, and again in the Warner recordings he made of both those pieces in 2018. Same pianist, same two works, five years and a career apart. It is the closest thing here to a controlled test of whether these numbers describe a person or an afternoon — and because there are two pieces, a change that is real should show up twice.

Read it across. One group of dials barely moves in either piece, and it is the group this post has been calling habits: how far he lifts the tune above the chord, how he weighs his hands against each other, how much his fingers overlap, how often his foot changes. The pedal changes per bar are identical to one decimal place in both pieces, five years apart.

Another group moves, and moves the same way in both pieces: he got slower, he spread his chords wider, he brought his hands closer together, he put the pedal down later, and he arched the phrases a little more. That is a coherent picture of what five years did — bigger, more deliberate chords, less of the old hands-apart habit — and the fact that the Brahms and the Schubert agree about it is what makes it worth saying. A third group (melody lead, beat sway, grain) moves in opposite directions in the two pieces, which is what a difference of reading rather than of hands looks like.

The archive holds a second case, longer and thinner. Jan Lisiecki played Ondine at the competition in 2008, aged thirteen, and recorded it for Deutsche Grammophon in 2022 — one piece, fourteen years.

The same split shows: what he lifts the tune by, how he weighs his hands, how much his fingers overlap are within a whisker of where they were; the tempo, the sway, the dynamic range and the pedal all moved. But notice that he moved opposite to Eric Lu on two of them — Lisiecki got faster and narrowed his chords where Lu got slower and widened his. Whatever fourteen years did to this pianist, it was not a general law of growing up. It was his.

Three recordings of two pianists prove nothing general. But this is the shape a real answer would have, and it is only visible because a competition happened to record two teenagers on pianos that write things down.

Sixty years of recordings

Ten recordings cannot show a trend over time; 397 dated professional recordings across 14 pieces can at least be asked. Within each piece, so that the piece's own level drops out, every dial was correlated with the recording year, and the correlations pooled across pieces. The oldest recordings here are from the 1950s. This is the one section the sensor recordings are kept out of: their dates all fall between 2002 and 2018 and they are all competitors at one event, so including them would load the recent end of every trend with a single cohort and let "what changed since 1950" quietly become "what a junior competition sounds like".

Michael Rector's finding — that tempo fluctuation across 127 recordings of a Chopin étude shows no trend from 1909 to 2016 — holds here too: arch depth (ρ = +0.08, p = 0.15) and overall tempo (−0.10) have not drifted since 1950, and beat sway shows at most a hint (+0.12, p = 0.02). What has moved is the pedal and the vertical. Recordings change the pedal less often the later they are made (−0.20, p < 0.01, negative in most pieces) — more blur, more resonance, a recorded-piano sound that has been getting wetter. And the chords are, if anything, coming further apart: roll width (+0.22), bass anticipation (+0.23) and melody lead (+0.21) all rise with the year, each with p < 0.01. Later recordings also make the tempo contrast between a piece's sections a little bigger (+0.14). Whatever died with the pre-war generation, it was not asynchrony as such. Two cautions: this says nothing about the dislocation of 1910, which is before the corpus begins; and older transfers are quieter and noisier, which a transcriber hears as fewer, more even notes, so some of the widening may be the microphone's.

What the words are, then

The five words survive this. They only lacked units, and what a corpus of 19 pieces adds is structure to the units. First, each word has a piece term and a pianist term, and only the second is style: the same dial can be mostly the music (voicing, rubato) or mostly the person (roll, float, dislocation, dynamic range). A compliment on a pianist's voicing is a compliment on a reading; a remark on their asynchrony is a remark on their hands. Second, there are habits and there are decisions, and an average measures only the first. Third, skill is not on these axes at all: home pianists differ from professionals in the scatter around each dial and in how far they dare to turn it, not in which dials they turn. Fourth, how much a piece leaves to the pianist is roughly constant, but where it leaves it is set by texture — and if anything the easy pieces are read more freely than the hard ones, which is the opposite of what "interpretation" usually implies. Fifth, the instrument is still ahead of the measurement: a piano with sensors under its pedals shows its players spending 26 % of the time half-pedalled, and no amount of listening to the recording recovers that. And the history, on 397 recordings, is that phrase-scale rubato has not changed since 1950, the pedal has been getting wetter, and the hands have been coming a little further apart rather than together. That is what the words are: five names for about thirty habits and a handful of decisions, some of them the pianist's, some the composer's, and most of them measurable now.

Play with it yourself

Every clip above was rendered in advance. This one is not: the whole piece, synthesised in your browser from the score and the consensus performance, with one slider per dial. Move them while it plays and the performance changes under your hands.

It knows all 19 pieces, and every one of the 624 recordings is in the Pianist menu — released recordings first, then the sensor recordings, then the home recordings. Choosing one sets all sixteen sliders to what this project measured in that recording, and says so if a dial could not be measured there. The buttons above the sliders are quicker: the ones marked as measured are the middle of a real group of recordings, and 🎲 picks a pianist at random, which is the fastest way to hear how far apart two people can be on the same notes. While the player is on screen, space plays and pauses and the arrow keys step a bar. The sliders' stops are set from the data too: each runs a little past the 2nd and 98th percentiles of everything anyone in the corpus actually did, so the far end of a slider is somebody's real playing plus a small margin, not a cartoon.

Caveats for the toy: the sound is a reduced sample set (pitches every minor third, two velocity layers) rather than the full soundfont used for the clips above, and the damper model is a simple one — a note rings until the pedal lifts, then fades. The dials are the real definitions, though, and the measured presets are this project's own numbers for those recordings. Two dials do not carry across between pieces unchanged: basic tempo is in each piece's own beat, and arch depth is mostly the gear change in a piece that has one, so both have their stops set per piece. Everything else — milliseconds, velocity units, per cent — means the same in every piece and shares one set of stops.

What to believe and what not to

• Onsets are trustworthy; releases and velocities less so. Transkun places note starts within about 10 ms. Releases are worse — nothing is audible when a key comes up under the pedal — and velocities depend on how the recording was mastered, which is why every loudness number here is a difference within one performance. The round-trip test against the sensor recordings puts numbers on all of this.
• Old and very quiet recordings are harder. Some 1950s transfers have the most missing notes, and some "pedal up" gaps may be the model failing to hear resonance under tape hiss.
• YouTube is the archive. Label uploads are studio recordings, but a release year is not always a recording year, and a handful of the home uploads are phone recordings of digital pianos, whose "pedal" is whatever the transcriber made of a sustain button. The soft-pedal dial is the least trustworthy of all: on covers it tracks the recording's brightness as much as the left foot.
• Two pieces are aligned to a score that is itself a performance. Avril 14th has no published score, so its reference is the composer's own piano-roll recording, quantised — fair to the other pianists and slightly flattering to him.
• Half-pedalling is invisible in the transcriptions, and the sensor recordings show it is not a small omission: they have their sustain pedal part-way down about 26 % of the time, a state a transcription has no symbol for. Touch quality — what the finger does before the hammer flies — is invisible in both. Continuous pedal-depth estimation and finger-key noise detection exist now; they would be the next two layers.
• The two kinds of recording are not interchangeable. The sensor recordings are more accurate but they are all competitors at one junior competition between 2002 and 2018, playing one make of piano. Where a number is quoted for a whole piece, they are in it alongside everyone else, so a piece with several of them (Schubert's Impromptu especially) leans a little more on young competition playing than the rest do.
• The corpus is uneven. Rosters run from 9 recordings to 67, and the home-recording sets from none to 16. The cross-piece sections say so where it matters. The dials are the durable part.

Appendix

Everything below is folded. Click a heading to open it.

1. Every recording, once — one passage from every recording of every piece
2. Three pianists, six pieces — Lang Lang, Ashkenazy and Biret, the same passage of each
3. The extremes, dial by dial — the two recordings at each end of every dial, per piece
4. Every number — the dials of every recording of every piece, and the full 32-descriptor table for one piece
5. Definitions — the 32 descriptors, precisely
6. Method and sources