Astra · page
OMR-NED 0.053
Joseph Haydn · String Quartet in D minor, Hob.III:76, Op.76 No.2
Every transcription from the high effort runs: scanned pages of printed string quartets, read by GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, Claude Fable 5.1, and Audiveris, then compared with OpenScore's transcription of the same pages. Lower OMR-NED is better.
These examples have the lowest OMR-NED among exported pages in each workflow. Select one to inspect every source crop and the rendered transcription.
Astra · page
Joseph Haydn · String Quartet in D minor, Hob.III:76, Op.76 No.2
Astra · crops
Antonín Dvořák · String Quartet No.10, Op.51
Fable · page
Joseph Haydn · String Quartet in D minor, Hob.III:76, Op.76 No.2
Fable · crops
Joseph Haydn · String Quartet in D minor, Hob.III:76, Op.76 No.2
Opus · page
Joseph Haydn · String Quartet in D minor, Hob.III:76, Op.76 No.2
Opus · crops
Antonín Dvořák · String Quartet No.10, Op.51
Sol · page
Franz Schubert · String Quartet in D minor, D.810, No.14 (“Death and the Maiden”)
Sol · crops
Antonín Dvořák · String Quartet No.10, Op.51
Audiveris · page
Edward Elgar · String Quartet, Op.83
Each cell opens that attempt below. Lower OMR-NED is better; longer blue bars mean closer notation. A page with no export still gets a score, because the scorer treats the missing transcription as empty. Astra is GPT-6 Astra, Sol is GPT-6.1 Sol, Opus is Claude Opus 5.5, and Fable is Claude Fable 5.1. Under each name, page means the model was given the full page, and crops means it was also given a separate image of each line of music.
Open an attempt to compare its source systems with the rendered transcription. Red means pitch, rhythm, or extra-note edits; amber means other note notation edits. The colors reflect comparison with OpenScore MusicXML, not direct image matching. Missing notes cannot be colored. The line images on the left are there for comparison. Only the runs marked crops were given them. Each line image has playback for its reference measure range. Snippets use ordinal measures and may drift after missing or inserted measures. Listen separately when simultaneous playback masks differences.
* MuseScore, the app that converts each answer to MusicXML, closed with an error right after saving this file. The file was checked as well-formed MusicXML with at least one part and was scored like the others.
39,400 reported tokens in 9 min 52 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
36,143 reported tokens in 8 min 18 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
26,688 reported tokens in 3 min 29 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
28,471 reported tokens in 4 min 28 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
42,242 reported tokens in 11 min 16 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
28,365 reported tokens in 4 min 20 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
1,723,894 reported tokens in 15 min 52 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
34,287 reported tokens in 7 min 51 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
45,781 reported tokens in 12 min 14 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
33,629 reported tokens in 6 min 36 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
29,776 reported tokens in 4 min 16 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
37,017 reported tokens in 7 min 47 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
37,765 reported tokens in 8 min 11 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
33,192 reported tokens in 5 min 53 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
35,731 reported tokens in 6 min 58 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
37,792 reported tokens in 8 min 21 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
81,878 reported tokens in 15 min 46 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
16,754 reported tokens in 1 min 53 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
19,879 reported tokens in 2 min 31 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
24,687 reported tokens in 3 min 28 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
24,496 reported tokens in 3 min 19 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
21,259 reported tokens in 2 min 38 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
68,462 reported tokens in 11 min 14 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
35,795 reported tokens in 4 min 54 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
86,340 reported tokens in 15 min 59 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
50,354 reported tokens in 8 min 17 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
22,831 reported tokens in 2 min 24 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
30,812 reported tokens in 3 min 47 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
42,819 reported tokens in 6 min 24 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
35,067 reported tokens in 4 min 58 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
154,340 reported tokens in 26 min 42 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
65,855 reported tokens in 10 min 29 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
55,869 reported tokens in 8 min 4 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
37,454 reported tokens in 5 min 15 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
20,034 reported tokens in 2 min 10 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
18,112 reported tokens in 2 min 4 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
27,691 reported tokens in 3 min 9 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
27,821 reported tokens in 3 min 27 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
42,445 reported tokens in 5 min 41 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
20,673 reported tokens in 2 min 18 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
53,786 reported tokens in 7 min 53 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
62,990 reported tokens in 9 min 5 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
25,468 reported tokens in 2 min 22 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
41,855 reported tokens in 5 min 15 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
39,622 reported tokens in 4 min 29 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
13,496 reported tokens in 1 min 23 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
29,762 reported tokens in 4 min 3 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
14,116 reported tokens in 1 min 34 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
3,010,659 reported tokens in 18 min 30 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
1,839,421 reported tokens in 16 min 4 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
31,857 reported tokens in 4 min 59 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
542,032 reported tokens in 7 min 12 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
819,867 reported tokens in 11 min 43 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
38,043 reported tokens in 7 min 35 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
2,656,404 reported tokens in 18 min 41 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
715,622 reported tokens in 7 min 25 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
3,013,017 reported tokens in 19 min 42 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
42,341 reported tokens in 9 min 13 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
478,723 reported tokens in 7 min 52 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
33,468 reported tokens in 4 min 49 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
50,231 reported tokens in 10 min 47 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
429,354 reported tokens in 7 min 37 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
1,911,545 reported tokens in 18 min 23 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
36,778 reported tokens in 5 min 57 s. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Diff tracks play only notes that sound different: reference notes the transcription missed or changed, and transcription notes with wrong pitch, accidental, duration, tie, or tuplet, or with no match. Matching notes become rests, so diff timing follows each full track.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Rendered transcription page unavailable.
Audiveris stopped before it could export a MusicXML file for this page.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Rendered transcription page unavailable.
Audiveris stopped before it could export a MusicXML file for this page.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
Rendered transcription page unavailable.
Audiveris stopped before it could export a MusicXML file for this page.
Token counts not applicable. Red marks show pitch, rhythm, or extra-note edits; amber marks show other note notation edits such as beaming. The comparison is to OpenScore MusicXML, so a colored note can have the correct pitch. Missing notes have no glyph to color. Playback clips use ordinal reference measures and may drift after missing measures.
OMR-NED by model and image condition; shorter bars are better. Bars use the same 0–1 scale.
Audiveris has one fixed baseline, with no effort or token count.
| Model | Condition | Medium Omr Ned | High Omr Ned | Ned Reduction Pct | Pages Improved | Pages Worsened | Pages Tied | Paired Export Pages | Paired Export Pages Improved | Medium Total Tokens | High Total Tokens |
|---|---|---|---|---|---|---|---|---|---|---|---|
| astra | page | 0.4184829700726521 | 0.29538423841465283 | 29.4 | 7 | 1 | 0 | 8 | 7 | 222213 | 1959490 |
| astra | crops | 0.453959984080466 | 0.3185366650334803 | 29.8 | 7 | 1 | 0 | 7 | 6 | 232613 | 290683 |
| fable | page | 0.6790634117516712 | 0.6585986839929902 | 3.0 | 5 | 3 | 0 | 7 | 4 | 143540 | 293210 |
| fable | crops | 0.7096817736982418 | 0.5263674187328755 | 25.8 | 7 | 1 | 0 | 8 | 7 | 141838 | 488418 |
| opus | page | 0.6094737565265183 | 0.54072696534235 | 11.3 | 7 | 1 | 0 | 8 | 7 | 126503 | 250099 |
| opus | crops | 0.5602175334733074 | 0.49638464657386133 | 11.4 | 5 | 3 | 0 | 8 | 5 | 142875 | 281095 |
| sol | page | 0.5485149823251512 | 0.1711252516070385 | 68.8 | 8 | 0 | 0 | 8 | 8 | 504636 | 9653905 |
| sol | crops | 0.5223428427164858 | 0.3212698801826372 | 38.5 | 7 | 1 | 0 | 7 | 6 | 559391 | 5995457 |
Paired export counts exclude medium pages without a MusicXML export.
Aggregate OMR-NED 0.763; 5 of 8 pages exported MusicXML. Audiveris used the same full-page scans, enlarged 3×, with OCR languages eng+deu+fra+ita. It is dedicated OMR software with a neural glyph classifier; it has no effort or token setting comparable to the model calls.
| Page | Composer | String quartet | OMR-NED | Outcome |
|---|---|---|---|---|
| page01 | Franz Schubert | String Quartet in D minor, D.810, No.14 (“Death and the Maiden”) | 0.831 | exported MusicXML |
| page02 | Joseph Haydn | String Quartet in D minor, Hob.III:76, Op.76 No.2 | 0.612 | exported MusicXML |
| page03 | Edward Elgar | String Quartet, Op.83 | 0.451 | exported MusicXML |
| page04 | Antonín Dvořák | String Quartet No.10, Op.51 | 0.591 | exported MusicXML |
| page05 | Ludwig van Beethoven | String Quartet No.1, Op.18 No.1 | 1.000 | no MusicXML export |
| page06 | Antonín Dvořák | String Quartet No.10, Op.51 | 1.000 | no MusicXML export |
| page07 | Franz Schubert | String Quartet in D minor, D.810, No.14 (“Death and the Maiden”) | 1.000 | no MusicXML export |
| page08 | Joseph Haydn | String Quartet in D minor, Hob.III:76, Op.76 No.2 | 0.648 | exported MusicXML |
Each answer is ABC text that MuseScore converts to a MusicXML file. This table counts how that last step ended for the 64 runs. Every run produced a file. A star means MuseScore saved the file and then closed with an error while shutting down. No star means it saved the file and exited normally.
| Outcome | Runs |
|---|---|
| exported MusicXML, MuseScore closed with an error after saving* | 49 |
| exported MusicXML, MuseScore exited normally | 15 |
* MuseScore, the app that converts each answer to MusicXML, closed with an error right after saving this file. The file was checked as well-formed MusicXML with at least one part and was scored like the others.
| Model | Images | Calls with usage | Processed input | Output | Median total / call | Total run time | Median run time / call |
|---|---|---|---|---|---|---|---|
| astra | crops | 8 | 173,462 | 117,221 | 36,374 | 60 min 16 s | 7 min 22 s |
| astra | page | 8 | 1,827,973 | 131,517 | 35,215 | 65 min 26 s | 8 min 4 s |
| fable | crops | 8 | 127,154 | 361,264 | 46,586 | 79 min 0 s | 7 min 20 s |
| fable | page | 8 | 79,522 | 213,688 | 24,592 | 45 min 43 s | 3 min 24 s |
| opus | crops | 8 | 83,434 | 197,661 | 34,692 | 36 min 4 s | 4 min 16 s |
| opus | page | 8 | 69,254 | 180,845 | 27,756 | 32 min 8 s | 3 min 18 s |
| sol | crops | 8 | 5,789,592 | 205,865 | 239,792 | 84 min 20 s | 8 min 32 s |
| sol | page | 8 | 9,430,163 | 223,742 | 767,744 | 92 min 9 s | 9 min 39 s |
Top three musicdiff edit costs per workflow, including the missing measures on pages with no export. All bars use the same 0–13,733 scale; longer means more edit cost.
Playback uses violin I, violin II, viola, and cello sounds. Diff tracks keep only the notes that sound different between reference and transcription, with matched notes turned into rests. Dynamics pp through mp play at mf; f and louder stay as written. Playback-only changes do not affect scoring.
OMR-NED is short for OMR normalized edit distance. It's an error rate built from the following sums:
Divide the first sum by the second and you get a number from zero to one. Zero means the two scores match. One means they have nothing in common. Accuracy is one minus that number:
Accuracy = 1 − OMR-NED = 1 − (I + D) / (Npred + Nref)
It counts symbols, not notes, so slurs, dynamics, and beams count along with pitches. This page shows OMR-NED itself, so lower is better. A page with no file is scored as an empty transcription, which gives 1.
For example, 229 edits with 1,068 symbols in the correct score and 1,072 in the transcription gives 229 ÷ (1,068 + 1,072) = 0.107. The score across all pages divides the sum of every page's edits by the sum of every page's symbols, so it is not the average of the page scores. Each run gets one score for the whole page, whether or not it was also given the separate line images. Playback-only changes to dynamics do not affect the score.
The page images are scans of printed editions from IMSLP, a library of public domain sheet music. They were downloaded from the OpenScore String Quartets page set published with LEGATO, which pairs each scan with OpenScore's transcription of the same page. The models and Audiveris were given the scans. OpenScore's transcription was used only as the answer key, and for the reference playback.
Each model read each page two ways: the full page alone (page), and the full page plus a separate image of each line of music (crops). The line images were cut from the same scan before the test, at the same size, to see whether giving each line its own image would help. The models did not cut anything up as part of this setting. This page shows the line images beside every transcription so you can compare one line at a time, but only the crops runs were given them.
The OpenAI models ran in Codex CLI 0.157.0 and the Claude models ran in Claude Code 2.1.282. Both versions were recorded on 2026-09-25, before the first run. The high effort runs took place from 2026-09-29 to 2026-10-07 (UTC), and the versions were not checked again for each run.
A run's time is the wall-clock time from starting the command to its exit. It does not include converting the answer to MusicXML. Many runs overlapped, so a time reflects the load at that moment as well as the model's speed. Audiveris run times were not recorded.
| Workflow | Visible input and output path |
|---|---|
| Astra (Codex CLI) | Page or page plus system crops → four-voice ABC → LEGATO cleanup → ABC to MusicXML → MuseScore 4 |
| Fable (Claude Code) | Same images, prompt, ABC output, and conversion path as Astra |
| Opus 5.5 (Claude Code) | Same images, prompt, ABC output, and conversion path as Astra |
| Sol (GPT-6.1 Sol, Codex CLI) | Same images, prompt, ABC output, Codex settings, and conversion path as Astra |
| Audiveris 5.11 | 3× enlarged full page → dedicated staff, symbol, text, and rhythm analysis → direct MusicXML export |
The model prompt forbids reference MusicXML, web, skills, and MCP tools. The tools were not the same for every model. Codex let the OpenAI models run commands, and Sol used that to crop and enlarge the scan on six of the eight pages. The Claude models could only open the image, so these results might make them look worse than they are. Audiveris includes a neural glyph classifier. The frontier models’ internal image-reading methods are not visible in these runs.