Harp sheet OMR allocates wrong voices

[SofNotenwart]

I have an excerpt from a piece for which I'm running Audiveris. On other instruments like violin etc, the recognition quality is quite good but for the Harp part, Audiveris doesn't allocate the correct voices to the music. It should be possible, I think at least, to reallocate the voices by selecting an element (note, rest) , right-click and under Chords select Preferred voice but it doesn't seem to work. The result in MS is that most measures have the wrong length.

Am I doing something wrong?

I'm not sure if this is a bug or that it cannot be done or something else.

I attach the original 3_Harfe_reduced.pdf and the 3_Harfe_reduced.omr file.

My environment is as follows:
Audiveris
Role: Optical Music Recognition
Version: 5.7.0:290525b
Container: file:/C:/Program%20Files/Audiveris/app/audiveris.jar
Home page: http://www.audiveris.org
Dev. project: https://github.com/Audiveris/audiveris
Software license: GNU Affero GPL v3
OCR engine: Tesseract OCR, version 5.5.1
OCR folder: C:\Users\tommy\AppData\Roaming\AudiverisLtd\audiveris\config\tessdata
Java vendor: Azul Systems, Inc.
Java version: 24.0.2
Java runtime: OpenJDK Runtime Environment (build 24.0.2+12)
Java VM: OpenJDK 64-Bit Server VM (build 24.0.2+12, mixed mode)
Operating System: Windows 11 10.0
Architecture: amd64

Audiveris was installed using the installer Audiveris-5.7.0-windowsConsole-x86_64.msi

[hbitteur]

I attach the original 3_Harfe_reduced.pdf and the 3_Harfe_reduced.omr file.

I didn't find any attachment.
I'd suggest you pack the .omr files into a zip archive, so that you can post them on this forum.

[SofNotenwart]

Here it is

HarpMusic.zip

[hbitteur]

@SofNotenwart

I fixed the problems in commit 7a84e06 on the development branch.
You can pull and build on your own.

Or you can wait until the next release (but I will be on vacation for the next 3 weeks in the US). I can quickly publish a draft release though.

I'm appending the resulting OMR file in the .zip file below.

23_Harfe_reduced.zip

FYI, I'm experimenting new techniques for a 6.x version of Audiveris, with more room for AI techniques. In your input score, many symbols got merged, and the current Audiveris version does not shine when trying to separate them.

I was curious to see how a new version could behave on this.
So, I ran a YOLO-based model on your input image, and here are the detected symbols I got in 15 seconds:

BINARY--1-ann.zip

[hbitteur]

Finally, I just published a 5.7.1 release so that you can download it and use the new version immediately.

I'll be out for the next 3 weeks, with access to my mail, but not to my development PC.