A measurement of piano playing

Repeat fidelity

When a passage comes back, how closely its timing repeats the first statement.

What it sounds like

The same passage — Pathétique II, bars 29–31 — from the recordings that sit lowest and highest on this measurement. Every clip is a note list replayed on one sampled grand at one gain, so the only difference you are hearing is the playing.

Least

Yundi Li, 2012 — bars 29–31, repeat fidelity 0.46 r16 s
Wilhelm Kempff, 1956 — bars 29–31, repeat fidelity 0.48 r12 s

Most

Stephen Kovacevich, 1972 — bars 29–31, repeat fidelity 0.96 r18 s
Claude Frank, 2011 — bars 29–31, repeat fidelity 0.96 r16 s

How it is calculated

Correlation between the beat-level log-tempo profiles of the two passages.

How much pianists differ on it

Every piece in the library, with the range across its recordings. The units mean the same thing in every row, except for basic tempo, which is in each piece’s own beat.

PieceRecordingsLowestMedianHighest
Rachmaninoff Prelude29-0.57-0.150.45
Nocturne Op. 48/127-0.070.260.54
Nocturne Op. 9/232-0.260.330.84
Brahms Op. 118/228-0.310.380.68
Schubert Op. 90/3430.070.590.96
Clair de lune270.520.720.91
Mozart K. 545370.070.761.00
Ravel Pavane330.010.800.94
Pathétique II470.460.810.96
Gymnopédie 1230.130.850.96
Schubert Op. 142/39-0.690.860.97
Fantaisie-Impromptu200.670.870.95
Träumerei310.600.890.97

How reliably it can be measured

A Disklavier capture — the pianist’s own key sensors, no microphone — was rendered to audio, transcribed as if it were a recording, and re-measured. The error the transcription step introduces, as a multiple of how far pianists actually differ on this dial, is 0.27. solid This dial survives transcription intact: where one recording sits on it can be argued from.