Dataset Viewer
Auto-converted to Parquet Duplicate
set
stringclasses
9 values
id
stringlengths
6
30
onset
float64
0.82
98.9
frame
float64
8.09
97.2
offset
float64
0.4
90.6
drums
float64
0
93.4
⌀
ref_notes
int64
142
24k
est_notes
int64
144
12.5k
ref_drums
int64
0
10.7k
est_drums
int64
0
7.33k
bach10
01-AchGottundHerr
91.61
95.05
86.71
null
142
144
0
0
bach10
02-AchLiebenChristen
93.1
94.27
90.62
null
283
282
0
0
bach10
03-ChristederdubistTagundLicht
89.26
94.65
84.85
null
183
180
0
0
bach10
04-ChristeDuBeistand
92.21
93.86
87.88
null
232
230
0
0
bach10
05-DieNacht
93.13
94.44
87.53
null
194
199
0
0
bach10
06-DieSonne
92.06
90.6
86.92
null
214
214
0
0
bach10
07-HerrGott
91.58
94.89
88.42
null
188
192
0
0
bach10
08-FuerDeinenThron
91.95
95.39
85.91
null
149
149
0
0
bach10
09-Jesus
90.12
94.37
87.79
null
170
174
0
0
bach10
10-NunBitten
94.04
94.6
90.21
null
234
236
0
0
dcs
DCS_LI_FullChoir_Take01
52.41
89.42
42.27
null
419
390
0
0
dcs
DCS_LI_FullChoir_Take02
62.1
91.67
52.32
null
419
399
0
0
dcs
DCS_LI_FullChoir_Take03
64.08
91.66
53.16
null
419
405
0
0
dcs
DCS_LI_QuartetA_Take01
56.23
83.43
44.02
null
419
367
0
0
dcs
DCS_LI_QuartetA_Take02
56.42
86.49
41.42
null
419
368
0
0
dcs
DCS_LI_QuartetA_Take03
39.36
73.57
27.13
null
419
333
0
0
dcs
DCS_LI_QuartetA_Take04
48.98
85.14
37.56
null
419
369
0
0
dcs
DCS_LI_QuartetA_Take05
46.06
85.59
36.3
null
419
380
0
0
dcs
DCS_LI_QuartetA_Take06
47.55
84.21
39.5
null
419
376
0
0
dcs
DCS_LI_QuartetB_Take01
72.39
91.08
56.93
null
419
396
0
0
dcs
DCS_LI_QuartetB_Take02
71.91
90.73
56.17
null
419
407
0
0
dcs
DCS_LI_QuartetB_Take03
63.11
89.03
48.54
null
419
405
0
0
dcs
DCS_LI_QuartetB_Take04
31.59
35.48
24.38
null
419
385
0
0
dcs
DCS_LI_QuartetB_Take05
70.23
90.72
55.16
null
419
404
0
0
dcs
DCS_TP_FullChoir_Take01
66.84
83.78
54.08
null
217
175
0
0
dcs
DCS_TP_FullChoir_Take02
47.21
82.53
34.01
null
217
177
0
0
dcs
DCS_TP_FullChoir_Take03
53.33
78.32
37.95
null
217
173
0
0
dcs
DCS_TP_FullChoir_Take04
49.49
65.17
46.94
null
217
175
0
0
dcs
DCS_TP_QuartetA_Take01
64.78
78.13
55.53
null
217
172
0
0
dcs
DCS_TP_QuartetA_Take02
52.97
84.34
45.05
null
217
187
0
0
phenicx
beethoven
71.43
84.52
29.95
null
2,657
3,360
0
0
phenicx
bruckner
64.63
82.39
31.91
null
2,527
2,938
0
0
phenicx
mahler
67.72
82
37.94
null
2,749
1,885
0
0
phenicx
mozart
81
84.26
60.11
null
2,657
2,563
0
0
rwc_p
RWC_P001
59.13
69.47
39.48
76.93
6,099
6,023
2,556
3,158
rwc_p
RWC_P002
43.07
79.12
27.91
26.75
5,704
3,082
683
2,465
rwc_p
RWC_P003
56.43
73.7
38.96
16.54
4,591
3,156
1,014
2,227
rwc_p
RWC_P004
50.58
76.87
31.62
37.43
4,803
2,907
3,370
2,882
rwc_p
RWC_P005
64.26
75.09
46.46
45.12
7,977
7,469
3,191
4,331
rwc_p
RWC_P006
43.7
66.16
29.49
61.24
10,800
4,323
2,226
2,033
rwc_p
RWC_P007
59.22
77.53
33.2
46.58
9,365
6,695
3,596
5,502
rwc_p
RWC_P008
70.24
79.75
52.32
65.61
7,219
5,438
1,185
1,531
rwc_p
RWC_P009
63.2
77.59
42.6
28.56
6,473
5,565
1,579
1,691
rwc_p
RWC_P010
47.91
70.6
31.85
52.59
6,394
5,976
2,155
2,154
rwc_p
RWC_P011
48.05
70.06
28.92
67.95
3,687
2,861
872
1,515
rwc_p
RWC_P012
50.26
77.51
30.39
76.44
6,534
4,489
1,673
1,812
rwc_p
RWC_P013
73.42
73.09
38.7
67.85
5,044
4,548
1,282
1,483
rwc_p
RWC_P014
55.35
69.79
28.62
56.88
4,289
5,528
2,336
2,105
rwc_p
RWC_P015
35.55
58.23
16.61
49.1
5,389
3,162
1,808
2,893
rwc_p
RWC_P016
45.4
67.64
15.58
62.24
13,275
6,088
5,030
4,411
rwc_p
RWC_P017
46.09
76.17
21.8
51.49
9,790
3,878
3,537
3,346
rwc_p
RWC_P018
57.5
77.42
37.65
26.08
8,920
5,024
1,146
3,010
rwc_p
RWC_P019
56.7
71.49
36.24
64.89
14,347
9,836
4,594
5,590
rwc_p
RWC_P020
72.85
81.45
42.45
53.17
3,711
3,408
2,389
1,907
rwc_p
RWC_P021
55.73
64.25
30.28
54.68
7,178
6,092
3,592
2,812
rwc_p
RWC_P022
54.91
69.18
35.14
46.6
6,603
4,284
1,764
2,524
rwc_p
RWC_P023
61.97
78.58
42.47
63.31
7,871
5,141
2,686
2,432
rwc_p
RWC_P024
70.89
79.73
48.59
39.68
5,663
6,093
2,537
3,894
rwc_p
RWC_P025
49.98
80.11
33.47
43.56
6,462
4,102
1,678
3,005
rwc_p
RWC_P026
38.13
48.94
15.64
72.96
5,436
3,130
2,069
1,955
rwc_p
RWC_P027
56.83
69.49
24.39
39.03
4,324
2,366
1,380
2,561
rwc_p
RWC_P028
44.5
76.85
31.56
16.12
11,860
4,580
777
2,474
rwc_p
RWC_P029
18.85
61.21
10.83
22.37
4,914
4,263
2,593
3,370
rwc_p
RWC_P030
64.91
76.24
42.22
74.6
4,344
3,159
1,705
2,062
rwc_p
RWC_P031
25.49
40.78
14.83
45.47
6,966
3,901
2,606
2,637
rwc_p
RWC_P032
63.47
71.31
51.18
55.06
4,655
3,806
2,190
2,409
rwc_p
RWC_P033
55.82
85.18
31.24
61.88
5,835
4,125
1,658
2,185
rwc_p
RWC_P034
72.09
78
45.03
27.83
2,956
2,978
959
1,607
rwc_p
RWC_P035
59.99
66.19
44.29
51.2
3,312
3,412
1,330
1,998
rwc_p
RWC_P036
55.43
68.91
35.93
43.91
8,387
7,032
1,970
2,644
rwc_p
RWC_P037
63.92
74.76
49.77
55.18
5,593
4,486
1,745
2,463
rwc_p
RWC_P038
68.06
82.33
48.94
60.4
5,773
5,625
2,017
2,513
rwc_p
RWC_P039
49.04
65.91
23.64
57.06
7,333
4,401
1,195
1,276
rwc_p
RWC_P040
65.55
77.53
33.32
44.61
6,017
5,459
3,284
2,679
rwc_p
RWC_P041
49.68
61.05
37.03
59.81
4,278
5,304
1,902
2,475
rwc_p
RWC_P042
52.69
72.37
31.01
58.03
8,959
5,518
6,542
4,638
rwc_p
RWC_P043
61.26
69.94
42.54
71.72
5,579
4,920
2,526
2,608
rwc_p
RWC_P044
60.97
82.26
34.25
37.4
6,331
5,674
6,203
4,311
rwc_p
RWC_P045
51.14
73.16
30.54
50.13
5,916
3,841
1,084
1,685
rwc_p
RWC_P046
61.43
77.66
42.87
44.31
7,169
5,213
2,714
2,522
rwc_p
RWC_P047
59.2
71.49
20.44
74.06
5,745
5,322
1,210
1,377
rwc_p
RWC_P048
47.65
67.2
25.38
31.48
8,189
4,335
3,875
2,052
rwc_p
RWC_P049
59.77
80.01
42.53
43.09
10,222
5,265
9,452
7,326
rwc_p
RWC_P050
57.66
71.9
34.21
53.7
5,654
3,489
2,068
2,822
rwc_p
RWC_P051
60.75
75.5
26.51
74.6
7,291
6,290
2,836
3,180
rwc_p
RWC_P052
67.55
74.77
40.45
49.14
8,015
5,830
2,661
2,459
rwc_p
RWC_P053
58.35
61.3
28.32
33.14
6,606
5,167
2,564
1,347
rwc_p
RWC_P054
62.34
81.46
25.3
58.92
6,158
6,065
2,325
2,397
rwc_p
RWC_P055
62.44
78.97
25.92
69.55
6,884
4,759
1,063
1,295
rwc_p
RWC_P056
32.84
64.95
15.6
27.17
5,291
7,040
5,163
3,441
rwc_p
RWC_P057
69.46
57.75
32.01
35.66
2,651
2,641
2,089
1,500
rwc_p
RWC_P058
70.52
76.47
44.19
49.31
7,923
6,261
4,397
5,180
rwc_p
RWC_P059
42.33
66.62
20.97
33.48
5,010
2,352
2,492
3,345
rwc_p
RWC_P060
50.26
66.86
33.75
37.55
4,545
2,685
5,618
3,095
rwc_p
RWC_P061
65.99
71.94
40.37
75.38
5,302
4,800
2,168
2,109
rwc_p
RWC_P062
57.87
71.16
29.49
65.86
2,590
3,662
1,323
1,316
rwc_p
RWC_P063
67.57
80.58
32.97
66.68
3,618
4,371
1,741
2,164
rwc_p
RWC_P064
62.23
79.98
37.56
66.2
6,132
4,879
866
1,276
rwc_p
RWC_P065
79.13
82.14
45.12
46.66
2,343
2,325
569
614
rwc_p
RWC_P066
78.38
78.27
38.43
68.7
3,074
3,073
872
1,058
End of preview. Expand in Data Studio

dettato-benchmark

Collection: dettato Code: kobimusic/dettato kobi.music

These are the benchmark outputs of KobiMusic's dettato, an audio to MIDI model. It has one MIDI file for each of the 513 recordings in the nine public sets from the model cards. It also has the score of each file and the scripts I used to make them and score them. If you score the MIDI files with scripts/score.py, they get 59.2 composite F1, which is the number on the cards.

What to expect. The MIDI files are the transcriber's notes plus the drum model's hits. That is the same setup as the cards. The audio is not here. The scripts download each set from its original source, so you could make the same files yourself. When I ran them again, the scores came out within 0.1 of the cards for each set, and the composite was the same.

What Is Here

path what
midi/<set>/<id>.mid the transcription of each recording (513 files)
results/per_recording.csv onset, frame, offset and drum F1 of each file, with the note counts
results/summary.json the scores of each set, the composite and its 95% interval
scripts/download.py downloads the nine sets from their original sources
scripts/build_sets.py turns the downloads into 16 kHz audio and reference notes
scripts/run_dettato.py runs the published models on every recording and writes the MIDI files
scripts/score.py scores a folder of MIDI files

Results

Each score is F1 x 100 for each recording, and then it is averaged for each set. I used the rules from the MuScriptor paper (Rouard et al. 2026). They use mir_eval and a 50-cent pitch tolerance.

  • onset: pitched notes with the onset within 50 ms. The instrument does not count.
  • frame: pitch-only piano rolls at 62.5 ms steps, scored cell by cell.
  • offset: an onset match, and the end within 50 ms or 20% of the note's length. The bigger one is used.
  • drums: drum hits with the onset within 50 ms on the exact GM key. Six of the sets have drums.
  • composite: the mean of the four. Each one is first averaged for the sets that have it. The 95% interval is a bootstrap on the recordings.
set recordings onset frame offset drums
Bach10 10 91.9 94.2 87.7 -
Dagstuhl ChoirSet 20 55.9 82.0 44.4 -
PHENICX-Anechoic 4 71.2 83.3 40.0 -
RWC Popular 100 57.1 71.9 33.1 50.6
RWC Classical 61 77.9 73.8 41.6 39.4
RWC Genre 102 51.0 65.7 29.3 36.3
RWC Jazz 50 63.5 67.5 36.0 36.3
RWC Royalty-Free 15 63.3 80.2 41.3 65.4
Slakh2100 test 151 73.5 82.3 49.4 54.4
all nine sets 513 67.3 77.9 44.8 47.1

The composite is 59.2 (95% interval 57.9-60.6). On the eight MuScriptor sets alone it is 58.4. MuScriptor 1.3B gets 45.1 on all nine sets and 44.6 on the eight, and YourMT3+ gets 41.1 on the eight. Those numbers are from the MuScriptor paper, except MuScriptor on Slakh2100, which I ran myself. The model card has them for each set.

Reproduce

You will need around 30 GB of disk. The downloads are 16 GB and the sets are 13 GB. A GPU is best for the models (a CPU will work too, but it is slow).

pip install -r scripts/requirements.txt
python scripts/download.py downloads/                 # about 17 GB, most of it RWC from Zenodo
python scripts/build_sets.py downloads/ sets/
python scripts/run_dettato.py sets/ out/              # about 4 h on an RTX 5060
python scripts/score.py sets/ out/transcriber --out results/

If download.py fails, you can run it again and it will only get the files it is missing. Zenodo is slow, at around 1 MB/s for each connection. So it gets the big RWC files in 4 parts at the same time. run_dettato.py also starts again from the first recording that is not done. It writes a second folder too, out/pipeline. Those files are what the full dettato command makes, with the evidence repairer and the grouper. They are not part of these results.

What should match. I made the sets again with these scripts, and they were the same as the sets I used for the cards. That includes the reference notes and the 16 kHz audio, sample for sample. Then I ran the models with the package on an RTX 5060. The first run was on an RTX 5090. On a GPU the models run in bf16, and there a different card or batch size could move a note by one frame. The cards used batches of 8, so run_dettato.py uses 8 as well (the dettato command uses 2). On a CPU the batch size makes no difference. In my rerun the transcriber wrote the same number of notes for each recording, and a few of them moved by one frame. Each set came out within 0.1 of the cards. The biggest change was the drums on RWC Classical, at 39.4 instead of 39.6. For the cards I made the drum features in a separate step, and the package makes them from the original file. The composite was 59.2 both times.

The Sets

set recordings hours source licence
Bach10 10 0.09 Bach10 v1.1 research use
Dagstuhl ChoirSet 20 0.68 Zenodo 4618287, V1.2.3 CC BY 4.0
PHENICX-Anechoic 4 0.18 Zenodo 840025 CC BY-NC-SA 3.0
RWC Popular 100 6.73 RWC 2.0, Zenodo 18656623 + rwc-annotations CC BY-NC 4.0
RWC Classical 61 5.45 same CC BY-NC 4.0
RWC Genre 102 6.97 same CC BY-NC 4.0
RWC Jazz 50 3.71 same CC BY-NC 4.0
RWC Royalty-Free 15 0.54 same CC BY-NC 4.0
Slakh2100 test 151 10.99 Slakh2100 mixes and YourMT3's stem MIDI, from Hugging Face CC BY 4.0

Some of the sets need a few choices to make the reference notes. They are all in build_sets.py:

  • Bach10: the notes come from the GTNotes.mat files. A note starts at its first frame and ends at its last frame.
  • Dagstuhl ChoirSet: the 20 takes that have a time-aligned score, with the room microphone.
  • PHENICX-Anechoic: the mix is the sum of every player's stem, since the set has no mix.
  • RWC: the aligned MIDI from rwc-annotations. Some vocal tracks are written 1 or 2 octaves above the sung notes. The file metadata_mapping_vocals.csv lists the octaves, so those tracks are lowered to the sung pitch.
  • Slakh2100: the stems that are rendered into the mix, with the bass tracks one octave lower. YourMT3 prepares this set the same way.

The 16 kHz audio is the input that the other models in the comparison got too. The drum model reads up to 20 kHz, so for the sets with drums it gets the original file.

The MIDI Files

  • The times are in seconds at 120 BPM, with 960 ticks per beat (about 0.5 ms).
  • The pitched notes are in the tracks named notes 1, notes 2 and so on. A MIDI track cannot have 2 overlapping notes on the same key. The transcriber can play a key twice at the same time, so I put the notes in as few tracks as that allows. You should read all of these tracks to get the pitched notes.
  • Drums is the drum track (channel 10) in the RWC and Slakh2100 files.
  • There are no instrument names. The transcriber only finds the notes.

Licence

The scripts are Apache-2.0, like the models. The MIDI files are transcriptions of the recordings in each set, so I would follow the licence of that set for them (see the table above). For RWC and PHENICX-Anechoic that means no commercial use.

Citation

@software{dettato2026,
  title  = {Dettato: multi-instrument audio-to-MIDI transcription},
  author = {Mills, Nathan},
  year   = {2026},
  url    = {https://github.com/kobimusic/dettato}
}
Downloads last month
40

Collection including kobimusic/dettato-benchmark