Scanner engines

OTS-HOMR

The scanner engine built for scores where several staves sound at once — string quartets, piano trios, voice and piano. It is a fork of HOMR that makes fewer mistakes on the same pages, reads the engraved details around the notes, and suggests the lyrics, written directions and dynamic marks it finds for you to review.

14.5% → 5.3%
String quartet scans
Symbols read wrongly, HOMR → OTS-HOMR; 792 staves, 9 scores
7.4% → 4.5%
Mixed repertoire
Symbols read wrongly, HOMR → OTS-HOMR; 3,422 staves, 632 scores
5.8% → 2.6%
Voice and piano scans
Symbols read wrongly, HOMR → OTS-HOMR; 300 staves, 13 scores

Full results · How we measured · Datasets and recipe

Which engine should I choose?

The scanner offers three. They differ in what they were built for.

EngineIn short
OTS-HOMRBuilt for ensemble scores. Reads staves together, reads stems, ties and slurs as printed, and suggests lyrics, directions and dynamics for review. This page.
HOMRThe open-source engine OTS-HOMR is forked from, maintained by its own authors.
TranscodaA different architecture entirely. Useful as a second opinion when a page reads badly.

There is no penalty for trying more than one. A page that defeats one engine often yields to another, and the comparison view shows you the two readings side by side.

How much more accurately it reads

We took staves from scores neither engine was built from, had each engine read them, and counted every symbol — note, rest, accidental, articulation, slur, position — that had to be added, removed or changed to turn the engine’s reading into the correct one. The error rate is that count divided by the number of symbols in the correct reading. Lower is better.

Test setStaves / scoresHOMR errorOTS-HOMR errorChange (95% interval)
String quartet scans792 / 914.47%5.25%−9.23 points (−12.37, −7.00)
Mixed repertoire3,422 / 6327.44%4.52%−2.92 points (−4.05, −2.03)
Voice and piano scans300 / 135.76%2.59%−3.16 points (−5.10, −1.54)

The improvement shows on every kind of symbol on every set. Wrong pitches or accidentals fall from 9.59% to 0.91% on quartet scans, from 7.32% to 4.07% on mixed repertoire, and from 5.93% to 1.26% on voice and piano. OTS-HOMR also reads time signatures HOMR cannot, such as five-beat and twelve-beat bars: 48 of 84 and 70 of 115 correct on mixed repertoire, none for HOMR.

Read these as a benchmark. The test staves were looked at during development, they are single staves rather than whole pages, and the correct answers for them were built by us and can contain errors of their own. The quartet figure rests on nine scores, a small slice of the 84 quartet scores we have corrected scans of.

Reading the staves together

HOMR reads a page one staff at a time. In a quartet the four staves play the same bars, so their barlines have to land in the same places; when they don’t, something was misread. OTS-HOMR uses that:

  • When a staff disagrees with the others about where bars fall, it tries the other likely readings of its doubtful rhythms and keeps the one that fits the other staves best. Over 200 pages this cut structural disagreements from 428 to 339, improved 81 of 899 systems, and made no page worse.
  • When most staves state one key or time signature and a few state another, it moves the few onto a value another staff actually read. It never invents one and never breaks a tie.
  • When one staff’s barlines drift from the rest, it tries other likely readings of that one bar, and keeps one only if its barlines then match the other staves exactly. Checking against the other staves is what makes this safe, and it is why this can’t be done on every staff from the start: each alternative means decoding again, and a single staff cannot tell which reading is right.

These apply to systems of at least three staves (four for the barline-drift repair), so they matter for quartets and trios, not for a single staff or a voice and piano pair.

It is deliberately cautious: music in several metres or keys at once is real, and a “fix” that overrides a genuine disagreement is worse than none.

Notation as printed

Standard recognition gives you pitches and rhythms and leaves the rest to a default. OTS-HOMR also writes out the stem directions, ties, slurs and horizontal spacing it reads from the page, so you can see and edit them as separate marks. Reading them is harder than reading the notes, so treat them as suggestions and check them as you would any other mark. Beaming is left to the editor, which groups notes by its own rules.

Lyrics, directions and dynamics

After a page is scanned, OTS-HOMR suggests the lyrics, written directions (cresc., rit., dolce…) and dynamic marks it found, each tied to a note, for you to check beside the original. Nothing is added to your score until you accept it, and marks it cannot attach to a note are still offered so you can choose one.

On clean song pages, the figures we measured:

  • About nine in ten suggested lyric placements were right (109 of 120, on pages the rules were not tuned on).
  • About half of the printed directions and dynamics on a page were offered, and about two in three of the offered ones landed in the right measure by a strict check against the encoded score.

Treat it as a first pass: it saves typing the lyrics and many of the marks, and you still check every one.

Worked examples

Two galleries. The first puts HOMR and OTS-HOMR side by side on single printed lines from songs and quartets that OTS-HOMR was not trained on. The lines are the clearest differences from a larger set, so the figures for the whole set are printed beside them. The second shows what the review-assisted lyrics, directions and dynamics add to a scanned page.

Each system is drawn once per reading, with a box on every note that differs; click a count to hide that kind of difference.

HOMR and OTS-HOMR, line by linepicked lines from songs and quartets OTS-HOMR was not trained on, beside the figures for the whole set
Lyrics and staff text, before and after reviewthe scan result, and the same page with every suggested lyric, direction and dynamic accepted

What it does not do

  • Every result is a draft. That is true of all music recognition; the review tools exist because of it.
  • It was measured on quartets, voice and piano, and mixed repertoire. We have not measured full orchestral or very dense scores.
  • It misses about half of the printed directions and dynamics. Loud marks such as f, ff and mf are found less reliably than soft ones, and wedge-shaped crescendo and diminuendo signs are not read.
  • Pages without music mislead it. A cover or an illustration can come back as an empty or nonsense measure, so leave those pages out of the scan.

Provenance and licensing

OTS-HOMR is derived from HOMR by Christian Liebhardt and contributors, and inherits its AGPL-3.0 licence. The recognition improvements described here are ours; the visual encoder and decoder they build on are HOMR’s.

Training data is built from OpenScore String Quartets, OpenScore Lieder and scans from IMSLP.

← Back to the scanner