Music Scanning Software

[tubatodd]

I'm a Linux guy at home and at work. I use open source software for just about everything. I've used MuseScore for music notation for several years and it is a nice application. The scan-to-notation "feature" (cloud hosted service) never works properly for me. Even with a pristine PDF, the results are lackluster at best. More often than not I find myself notating 1 note at a time from scratch. Ugh!

What is everyone else using for scan-to-notation software? How accurate is it?

[gocsick]

I did a deep dive on this a few months back... the short answer is Musescore uses the Audiveris library which is widely regarded as the state the art programming for OCR of music... dyo there really isn't anything better at the moment.


It turns out to be a really really challenging computer science problem. It combines multiple hard problems from computer vision, pattern recognition, and musical semantics.

Symbols are complex and overlapping: Unlike printed text, music has nested symbols (e.g., beams, tuplets, slurs, articulations) that often intersect or overlap.

Two-dimensional structure: Music encodes pitch vertically and time horizontally, so interpreting layout is nontrivial.

Dense information: A single staff system might pack dozens of elements that must all be correctly interpreted and related.


PDFs Are Not Created Equal

Scanned PDFs are images—often fuzzy, distorted, or poorly lit.

Digital PDFs (exported from notation software) often contain vector data, but not semantic structure (e.g., the fact that a symbol is a staccato dot tied to a specific note).

So even in the best-case scenario (digital PDFs), there’s no guarantee the MusicXML structure can be recovered without inferring intent.


Ambiguity in Music Notation

Contextual interpretation: A symbol like a dot can mean a staccato or a dotted rhythm—depending on placement.

Polyphonic music: Identifying voices correctly (e.g., SATB on a grand staff) requires musical understanding, not just visual parsing.

Implied meaning: Notation often leaves information implied (e.g., omitted time signatures, unlabelled repeats, or accidentals carried across measures).

These require more than symbol recognition—they require music understanding.


OMR Software is Fragile

Most current tools (like SmartScore, PhotoScore, Audiveris, or PlayScore) are:

Rule-based or Machine learning -hybrid, not fully learned end-to-end.

Trained on limited datasets—they struggle with handwritten, unusual, or historical scores.

Not user-friendly—they may require tons of manual correction, and Musescore doesn’t make fixing errors easy.


Even state-of-the-art tools like Audiveris (open-source) or commercial tools like ScanScore typically require significant manual editing.


Basically new approaches are utilizing AI told which requires training on paired parts... original score and music XML... The challenge is getting the thousands of scores necessary to train such a model.

[tubatodd]

Thanks @gocsick. I appreciate the complete detailed reply.

The band leader of the big band I play bass in had asked me about scanning a score in order to change the key. I had no advice for her.

There have been many times I've been given a part that needed editing or completely re-notated because the chart was tough to read. I attempted to make a flat black and white image with as few artifacts as possible. I've even removed tempo, dynamics and other text contexts. It always generated something unusable or the MuseScore service would just refuse to produce a document. Are people actually able to get some value out of this? Am I doing something wrong with my input file?

It's a shame this isn't as "simple" as TWAIN.

[Richard III]

I've been using Sibelius for years. I should also say it is the subscription version where I pay $20 a month. The advantage of that is that my version is always the current one and frequently updated. If the PDF is quality and clean, it scans almost without errors and makes changes so easy. I'm primarily a trumpet player and play everything in treble clef. That means if I'm playing euphonium, trombone or tuba, the parts are transposed into treble clef to make it easy for me. I also use it to alter parts for a variety of instruments for my small jazz band when I can't find a player for the part on the original instrument. For instance, my bass player really wanted to play a solo originally written for trumpet. With a couple of key strokes, he has the part in bass clef and plays it without a thought.

All that being said, if the original is not clean and clear, nothing will scan well. There is no program that will give you a result that is not full of errors that take so long to correct, you might as well have done it by hand.

[Colby Fahrenbacher]

I've never had any luck with the MuseScore scan-to-notation feature either. I try a PDF every once in a while to see if they've made any progress, but the result (if there is one) is such a mess that it would be safer and easier to just re-engrave it from scratch.

It's reassuring to hear that the Sibelius equivalent seems to be successful. Does Dorico have this feature yet?

[gocsick]

Sibelius uses Photoscore for it's OMR engine.. Photoscore is Commercial and costs $250 to buy the full version...where Audioveris is open source and freely available.

I would not be surprised if it performs better... on the other hand I've seen plenty of open source tools that are miles ahead of their commercial counterparts.

[GC]

I've used ScanScore for Mac to scan treble clef brass band Bb Bass parts and put them into bass clef. It's not an easy process and there are tons of errors to correct, but it's better than doing it by hand. Its editing and typesetting tools are pretty good. It's definitely time consuming, but gets decent results, and it certainly beats writing notes by hand with a buttload of ledger lines below the staff.

Has anyone used Dorico, the replacement for Finale?