Figure 1
Overview of the Beethoven Symphony Excerpt Dataset. The dataset contains 20 short symbolic score excerpts from Beethoven’s symphonies, each paired with five audio versions from different performances, including one synthetic version. Shown here as an example is the first excerpt only. All excerpts feature precisely aligned note‑level annotations with pitch and instrument labels, offering a controlled, well‑verified benchmark for testing orchestral automatic music transcription systems.
Table 1
Overview of the excerpts in the BSED with their corresponding audio versions. The measure numbers are inclusive and refer to the Urtext edition (Del Mar, 2020), of which we only use full measures (except for the anacrusis in BSED‑03). The five audio versions comprise four real performances (BSED‑Real) as well as a synthetic rendition (BSED‑Sytnh).
| BSED‑ID | MovementID | Measures | # Notes | Avg. Dur. [s] | BSED‑Real | BSED‑Synth | |||
|---|---|---|---|---|---|---|---|---|---|
| Split 1 | Split 2 | Split 3 | Split 4 | Split 5 | |||||
| BSED‑01 | Op021‑01 | 1–4 | 159 | 26.8 | Karajan1963 | Ansermet1956 | Keilberth1958 | Blomstedt1976 | Synth |
| BSED‑02 | Op021‑02 | 6–16 | 256 | 21.8 | Krips1962 | Paray1959 | Nowlin2019 | Blomstedt1976 | Synth |
| BSED‑03 | Op036‑01 | 0–35 | 73 | 17.4 | Schuricht1948 | Leibowitz1961 | Schubert2005 | Blomstedt1978 | Synth |
| BSED‑04 | Op036‑01 | 35–48 | 556 | 19.5 | Paray1959 | Karajan1963 | Steinberg1964 | Blomstedt1978 | Synth |
| BSED‑05 | Op036‑03 | 1–26 | 401 | 17.0 | Leibowitz1961 | Paray1959 | Schubert2005 | Blomstedt1978 | Synth |
| BSED‑06 | Op055‑01 | 1–17 | 319 | 21.5 | Pfitzner1929 | Ormandy1961 | Krips1958 | Blomstedt1979 | Synth |
| BSED‑07 | Op055‑02 | 143–148 | 320 | 23.1 | Markevitch1957 | Klemperer1961 | CzechNat20126 | Blomstedt1979 | Synth |
| BSED‑08 | Op055‑03 | 231–2647 | 234 | 21.8 | Steinberg1955 | Klemperer1961 | Steinberg1963 | Blomstedt1979 | Synth |
| BSED‑09 | Op060‑01 | 1–6 | 72 | 28.5 | Weingartner1952 | Karajan1962 | Keilberth1959 | Blomstedt1979 | Synth |
| BSED‑10 | Op060‑02 | 98–101 | 279 | 21.6 | Weingartner1952 | Karajan1955 | Keilberth1959 | Blomstedt1979 | Synth |
| BSED‑11 | Op067‑01 | 1–21 | 244 | 21.2 | Furtwaengler1943 | Bernstein1963 | Karajan1962 | Blomstedt1977 | Synth |
| BSED‑12 | Op067‑04 | 110–120 | 367 | 16.7 | Melgelberg1937 | Schuricht1949 | Jochum1955 | Blomstedt1977 | Synth |
| BSED‑13 | Op068‑01 | 1–26 | 300 | 27.7 | Hurst1963 | Lehmann1956 | Toscanini1956 | Blomstedt1977 | Synth |
| BSED‑14 | Op068‑05 | 194–201 | 497 | 16.9 | Kubelik1960 | Hurst1963 | Cluytens1960 | Blomstedt1977 | Synth |
| BSED‑15 | Op092‑02 | 1–18 | 171 | 30.2 | Monteux1964 | Bernstein1960 | Karajan1962 | Blomstedt1975 | Synth |
| BSED‑16 | Op092‑04 | 31–49 | 525 | 16.5 | Monteux1953 | Ormandy1945 | Krip1961 | Blomstedt1975 | Synth |
| BSED‑17 | Op093‑01 | 1–20 | 769 | 22.2 | Leibowitz1961 | Ansermet1956 | Keilberth1958 | Blomstedt1978 | Synth |
| BSED‑18 | Op093‑02 | 30–35 | 322 | 17.8 | Leibowitz1961 | Ansermet1956 | Keilberth1958 | Blomstedt1978 | Synth |
| BSED‑19 | Op125‑01 | 374–385 | 376 | 21.7 | Klemperer1956 | Furtwaengler1955 | Leibowitz1961 | Blomstedt1980 | Synth |
| BSED‑20 | Op125‑03 | 3–7 | 102 | 27.3 | Furtwaengler1955 | Fricsay1958 | Leibowitz1961 | Blomstedt1980 | Synth |
| Total: | 6342 | 37:14 min | 7:26 min | 7:36 min | 6:58 min | 7:46 min | 7:29 min | ||
Figure 2
Total number of notes on the symbolic level per instrument in the BSED.
Figure 3
Number of notes on the symbolic level of each BSED excerpt grouped by instrument. Instruments are ordered first by instrument family and then by their total number of notes across all excerpts within each family.
Figure 4
Number of notes on the symbolic level per pitch in the BSED scores grouped by instrument.
Figure 5
Sequence‑level versus note‑level alignment for a chord from BSED‑01. All notes share the same musical onset time in the score. The spectrogram is shown at the transcription model’s time resolution (hop size: 32 ms). The dashed black line indicates the common physical onset time assigned by sequence‑level alignment, whereas the red boxes mark the individual physical onset times obtained by note‑level alignment.
Table 2
Overview of the collected audio of the Beethoven Symphonies Dataset for each symphony and movement. The fourth movement of Symphony No. 9 is reported separately in brackets because of its distinct instrumentation. In total, the dataset comprises 411 + [4] movement recordings.
| Opus | Name | # Mvmt. 1 | # Mvmt. 2 | # Mvmt. 3 | # Mvmt. 4 | # Mvmt. 5 | Total Dur. [hh:mm] |
|---|---|---|---|---|---|---|---|
| Op. 21 | Symphony No. 1 in C major | 11 | 11 | 11 | 11 | — | 04:34 |
| Op. 36 | Symphony No. 2 in D major | 8 | 8 | 8 | 8 | — | 04:19 |
| Op. 55 | Symphony No. 3 in E major | 19 | 18 | 18 | 17 | — | 14:33 |
| Op. 60 | Symphony No. 4 in B major | 5 | 5 | 5 | 5 | — | 02:45 |
| Op. 67 | Symphony No. 5 in C minor | 16 | 15 | 15 | 13 | — | 07:50 |
| Op. 68 | Symphony No. 6 in F major | 18 | 19 | 16 | 18 | 18 | 12:41 |
| Op. 92 | Symphony No. 7 in A major | 17 | 17 | 16 | 17 | — | 10:05 |
| Op. 93 | Symphony No. 8 in F major | 4 | 4 | 4 | 4 | — | 01:43 |
| Op. 125 | Symphony No. 9 in D minor | 4 | 4 | 4 | — [4] | — | 02:59 + [01:33] |
| 102 | 101 | 97 | 93 + [4] | 18 | 61:28 + [01:33] |
Figure 6
Warping paths between audio and score for two different unfoldings of the same piece: (a) Structurally equivalent pair (b) Structurally non‑equivalent pair.
Table 3
Overview of different data split strategies for the BSD, when testing on BSED. Each split defines how recordings are excluded, with the corresponding number of training recordings and test excerpts shown. ‘+ [4]’ indicates recordings of Op. 125‑04 separately as in Table 2.
| Data Split | # Train Rec. | # Test Excerpts |
|---|---|---|
| Movement Split | 193 + [4] | 100 |
| Version Split—Whole Symphony | 176 + [1] | 100 |
| Version Split—Single Movement | 351 + [4] | 100 |
| Neither Split | 85 | 100 |
| Version Split Reduced BSED | 411 + [4] | 40 |
Table 4
Note‑level precision (P), recall (R), and F1‑score (F1) (%) (onset only) for different evaluation sets. Rows represent transcribers trained on BSD with different alignment methods, and columns correspond to evaluation sets with the same three alignment methods applied to BSED‑Real, as well as BSED‑Synth and PHENICX dataset (which do not rely on the alignment strategies). The best scores for BSED‑Synth and PHENICX are boldfaced.
| BSED‑Real | BSED‑Synth | PHENICX | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| P | R | F1 | P | R | F1 | P | R | F1 | P | R | F1 | P | R | F | |
| 72.2 | 34.2 | 43.8 | 57.9 | 39.9 | 37.7 | 73.1 | 36.1 | 45.9 | 77.6 | 26.9 | 36.5 | 79.8 | 36.0 | 49.4 | |
| 50.0 | 38.8 | 43.2 | 81.7 | 61.7 | 69.2 | 84.9 | 63.9 | 71.6 | 88.6 | 60.2 | 70.2 | 79.8 | 49.0 | 60.4 | |
| 36.5 | 38.2 | 37.1 | 66.6 | 69.2 | 67.3 | 70.5 | 73.3 | 71.3 | 73.8 | 69.0 | 70.6 | 74.1 | 59.4 | 65.7 | |
Table 5
Note‑level transcription results for the BSED and PHENICX test sets in % with an onset tolerance of 50 ms. For BSED, we use the annotations.
| Test Set | Transcriber | Note (onset only) | ||
|---|---|---|---|---|
| P | R | F1 | ||
| BSED | BasicPitch | 32.3 | 23.5 | 24.9 |
| YMT3+ | 47.4 | 39.9 | 42.3 | |
| 67.6 | 57.3 | 60.6 | ||
| 71.1 | 72.5 | 71.1 | ||
| 73.1 | 70.4 | 70.8 | ||
| PHENICX | BasicPitch | 54.1 | 34.7 | 41.8 |
| YMT3+ | 70.3 | 48.3 | 57.0 | |
| 77.5 | 56.3 | 64.9 | ||
| 74.1 | 59.4 | 65.7 | ||
| 78.5 | 62.0 | 69.0 | ||
Figure 7
Mean F1‑scores for note‑level transcription (onset only) of across evaluation onset tolerances for BSED‑Synth, BSED‑Real with sequence‑aligned () and note‑level–aligned () annotations, and PHENICX.