Table 1
Overview of datasets used for our distant listening approach, totaling over 600 hours of music.
| Dataset | Pieces | Performances | Duration | MIDI | Repertoire |
|---|---|---|---|---|---|
| Magaloff/Chopin (Flossmann et al., 2010) | 155 | 155 | 8 h 17 m | recorded | Essentially all of Chopin’s solo piano work |
| Zeilinger/Beethoven (Cancino‑Chacón et al., 2017) | 30 | 30 | 2 h 40 m | recorded | 9 piano sonatas: Ops. 2, 13, 14, 26, and 27 (early); Ops. 31 and 53 (middle); and Ops. 109 and 110 (late) |
| Batik‑plays‑Mozart (Hu and Widmer, 2023) | 36 | 36 | 3 h 46 m | recorded | 12 piano sonatas: KV279–284, KV330–333, KV457, and KV533 |
| Vienna4×22 (Goebl, 1999) | 4 | 88 | 2 h 18 m | recorded | Excerpts from 4 pieces by Chopin, Mozart, and Schubert |
| (n)ASAP (Peter et al., 2023) | 235 | 1067 | 94 h 30 m | recorded | Common Practice Period piano solo work by 15 composers |
| ATEPP (aligned subset) (Zhang et al., 2022) | 320 | 5094 | 524 h 48 m | transcribed | Solo piano work by 25 composers, ranging from Baroque to the Modern era |
| Total | 780 | 6330 | 636 h 10 m |
Figure 1
Tempo histogram over median normalized tempi across all (1326) performances of (80) pieces by Chopin in the ATEPP dataset, with the resulting tertile‑based tempo grouping. Red dashed lines indicate the mean positions of the first and second tertile boundaries, while the shaded regions represent the corresponding standard deviations.
Table 2
Average correlations between performances of pieces by Beethoven and Chopin per dataset. For these two composers, mean within‑group correlations (main diagonal) were slightly higher than the mean between‑group correlations (lower triangle).
| Composer | Dataset | Tempo | Fast | Medium | Slow |
|---|---|---|---|---|---|
| Beethoven | ATEPP | fast | 0.22 | – | – |
| medium | 0.26 | 0.30 | – | ||
| slow | 0.26 | 0.33 | 0.39 | ||
| (n)ASAP | fast | 0.45 | – | – | |
| medium | 0.49 | 0.51 | – | ||
| slow | 0.49 | 0.54 | 0.60 | ||
| Chopin | ATEPP | fast | 0.35 | – | – |
| medium | 0.33 | 0.35 | – | ||
| slow | 0.32 | 0.33 | 0.43 | ||
| (n)ASAP | fast | 0.29 | – | – | |
| medium | 0.32 | 0.35 | – | ||
| slow | 0.33 | 0.35 | 0.38 | ||
| Vienna4×22 | fast | 0.80 | – | – | |
| medium | 0.72 | 0.81 | – | ||
| slow | 0.80 | 0.80 | 0.81 |
Table 3
Mean and standard deviation values of the proportion of pairwise correlation comparisons per piece where the between‑group correlation (slow vs. fast tempo) is significantly lower than the within‑group correlation (), reported separately for the slow and fast groups, each serving in turn as the within‑group benchmark. Columns pc and pfc indicate piece and performance count by that composer in the respective dataset.
| Composer | Dataset | pc | pfc | Within = Slow | Within = Fast | ||
|---|---|---|---|---|---|---|---|
| %μ | %σ | %μ | %σ | ||||
| Beethoven | ATEPP | 25 | 770 | 53.95 | 14.20 | 26.55 | 10.40 |
| (n)ASAP | 28 | 212 | 58.79 | 25.39 | 25.30 | 29.20 | |
| Chopin | ATEPP | 80 | 1326 | 45.46 | 27.71 | 36.80 | 27.59 |
| (n)ASAP | 26 | 271 | 39.42 | 27.18 | 15.50 | 11.09 | |
| Vienna4×22 | 2 | 44 | 21.79 | 8.82 | 11.90 | 11.17 | |
Table 4
Excerpt of the analysis of variance results, in terms of absolute and relative amounts of performances of works by Beethoven and Chopin for which a significant effect (P < 0.05) for Metre/Tempo was found, shown separately by dataset. Both metrical strength and tempo group significantly affect timing patterns, but, in most pieces, their influence is largely additive rather than interactive.
| Composer | Dataset | Metrical Strength | Tempo Group | Interaction | |||
|---|---|---|---|---|---|---|---|
| % | % | % | |||||
| Beethoven | ATEPP | 23 | 92.00 | 25 | 100.00 | 2 | 8.00 |
| (n)ASAP | 19 | 65.52 | 21 | 72.41 | 5 | 17.24 | |
| Chopin | ATEPP | 54 | 67.50 | 70 | 87.50 | 14 | 17.50 |
| (n)ASAP | 24 | 92.31 | 23 | 88.46 | 2 | 7.69 | |
| Vienna4×22 | 2 | 100.00 | 2 | 100.00 | 0 | 0.00 | |
Figure 2
Illustration of key overlap time and inter‑onset interval for legato (top) and staccato (bottom) scenarios. Absolute timestamps of the score events are shown in blue.
Figure 3
Distribution of sustain pedal values (CC 64) aggregated across all pieces (and pianists) for each dataset. The shaded region marks the intermediate range (control values 50–100) used to locate the modal accumulation corresponding to the physical damping point. The solid vertical line shows the estimated sound‑off threshold (Vienna422: 76, Batik: 75, Magaloff: 60); the dashed line marks the conventional default of 64.
Figure 4
Key overlap ratio distribution of legato‑ and staccato‑marked notes in different datasets. The vertical axis shows a mirrored kernel density estimate histogram for visual comparison between articulation groups. Legato values indicate mirrored density and do not represent negative probabilities.
Figure 5
Boxplots of the key overlap ratio for all note transitions, grouped by interval consonance level across datasets. The unison (same pitch) transitions are grouped in the left (upward) side.
Figure 6
Relationship plot between the octave register and key overlap ratio across datasets. The shaded area represents the 95% confidence interval, interpolated between discrete points for each octave.
Figure 7
Mean root mean square (RMS) asynchrony (in ms; lower axis and colored bars) and polyphonic density (expressed as a proportion; upper axis and black dashed line) by composer and metrical position: downbeat (top), beat (middle), and offbeat (bottom). Polyphonic density indicates the proportion of events at each metrical position containing 3+ simultaneous notes. RMS error bars show 95% confidence intervals across performers for mixed‑performer datasets or alternatively across pieces for single‑performer datasets. Polyphonic density error bars show standard deviation across pieces.
Figure 8
Melody lead time (ms) as a function of MIDI velocity difference between melody and accompaniment (binned at four‑unit intervals), shown for right hand only (top) and both hands (bottom), at downbeat, beat, and offbeat positions. Violin plots show the distribution per velocity bin across three corpora (ATEPP, (n)ASAP, and Vienna4×22); dashed lines are dataset‑specific regression fits on unbinned data. Melody lead increases with velocity difference, most strongly for the right hand at downbeat positions.
Figure 9
Comparison between ATEPP (transcribed MIDI) and (n)ASAP (recorded MIDI) in terms of timing features discussed in previous experiments, computed on a subset of 125 pieces present in both datasets: score‑normalized, log‑scaled inter‑onset interval values (left, discussed in Section 5) and overall root mean square asynchrony (right, see Section 7), both as a function of estimated tempo (beats per minute). Each point corresponds to one performance‑wise mean.