Sharp, rests and chordnames not recognized from fakebook

[fxpar]

Hi,

I am trying to transcribe the attached fakebook. But the sharp and the rests (first page) are almost never recognized, neither are the chordnames (page 2).

Following the advice given similar questions, I have tried:

• Changed the fonts in the book parameters (for the whole book) Actually, I created a test pdf with six times the same sheet to test each font. No luck.

Do you have any other settings that might have an effect?

Thanks for your help.

SharpSilenceChordnames.pdf

[hbitteur]

Trying your example, I got:

[SharpSilenceChordnames#1] Staff#1 no recognized header clef.
[SharpSilenceChordnames#1] Staff#2 no recognized header clef.
[SharpSilenceChordnames#1] Staff#3 no recognized header clef.
[SharpSilenceChordnames#1] Staff#4 no recognized header clef.
[SharpSilenceChordnames#1] Staff#5 no recognized header clef.
[SharpSilenceChordnames#1] Staff#6 no recognized header clef.
[SharpSilenceChordnames#1] Staff#7 no recognized header clef.
[SharpSilenceChordnames#1] Staff#8 no recognized header clef.
[SharpSilenceChordnames#1] Staff#9 no recognized header clef.

Indeed each staff line should begin with a clef located within the staff width. Otherwise Audiveris barks!

Now, to answer your question, I have no solution out-of-the-box.

Changing the music font is efficient for the recognition of head notes (since their recognition is based on template matching).
But for all the other fixed-shape symbols, like the sharp, rest, flat, ... the engine uses a small neural network, trained on a population of representative samples. And "your" symbols are not well known enough.

I'm currently working on a new and more powerfull neural network, which should lead to better results.
Meanwhile, could you share with us several scores with similar (but not identical) symbol shapes, so that we can train on them?

[fxpar]

Hi / Bonjour Hervé,

First thank you very much for your answer. It is really kind of you to try my example. I am looking forward to trying the new neural network (hopefully, it will not be too costly on GPU).

As you noticed, the scores are really clean.
Most of the jazz fakebooks are from this collection on archive.org: https://archive.org/details/fakebooks

If you want the symbols of the "same kind" but not the same exact font, you might try thoses (but they are a lot dirtier than the one I sent you)

This one seems "a bit clean": https://archive.org/download/jazz-books-realbook/Jamey%20Abersold/Vol%2031%20-%20Bossa%20Novas%20%28Bb%20%26%20Eb%29_text.pdf

Maybe dirtier versions: https://archive.org/details/500TunesRealBookMissed/page/n111/mode/2up
Or this one: https://archive.org/details/fakebook_the-real-book---fifth-edition---f---bass-clef/page/n117/mode/2up

Hope this helps.

[hbitteur]

I am looking forward to trying the new neural network (hopefully, it will not be too costly on GPU).

FYI, we could accept a GPU requirement, but only at training time.
The goal for the end user is still to be able to run the application on a standard PC using its CPU at inference time.

[hbitteur]

@fxpar

I'm using your examples to feed the new implementation of the glyph classifier. So far, things are going well.

Meanwhile, I noticed that all these scores seem to use a similar music font and a similar text font.
Do you know the names of these fonts? And where we could get them?
If I could insert them into Audiveris, we would get significantly better results.

[fxpar]

Hi Hervé,

Gemini suggested the following fonts (MuseJazz, LilyJazz, JazzFont - the Finale original font, Inkpen or Inkpen2) as the most used font in realbooks or fakebooks.
I don't know which one is used, sorry.
Hope this helps, though.

[hbitteur]

@fxpar
The major goal is the music (rather than text) font.

See this image below (extracted for the Bossa Novas pdf file):

The 8th rest and the flag will be recognized by the glyph classifier, when duly trained on these samples.

But the problem is with the head shapes (here a void head and a black head), because Audiveris uses a template matching technique to recognize them. And there is no training possible, we must provide a suitable music font, for templates to match correctly.
Notice how "thin" the black head is!

I'm still investigating music fonts on the web, with no great success so far.

[hbitteur]

I'm having problems exporting the transcribed score to MuseScore (or any other music software able to read some MusicXML data)).
Here is the beginning of "Meditation" score, in Bossa Novas:

Here is Audiveris transcription (there was no attempt to deal with chord names)

And here is the MuseScore display:

I think you can understand what happened.
A staff is supposed to begin with the sequence of clef + (optional) key signature + (optional) time signature

The problem is the input score does not respect these rules in staves 2 and 3.
No starting clef, no key sig, no time sig.
Consequently, MuseScore considers that staff 2 begins with a new (empty) key sig (and warns the reader at the very end of staff 1)
Similarly, at then end of staff 3, just before the repeat sign, there is a warning about shifting from an empty key sig to a 2-sharp key sig.

These are the rules used in music common notation. And the input scores don't comply with these rules.

[hbitteur]

Since many of your examples, if not all, are out of rule, is there a way to process them anyway, probably via some specific processing.

What comes to mind is to consider that:

• A staff with a starting clef is the beginning of a "sequence of staves" (starting from scratch)
• A staff without a starting clef is just the continuation of the preceding staff (same clef/key/time)

Would this make sense?

What puzzles me is the following case (using the same Meditation example):

1. First staff starting with treble clef, 2 sharps, 4/4 time sig
2. follows staff above
3. follows staff above
4. follows staff above
5. follows staff above
6. follows staff above
7. follows staff above
8. Staff starting with treble clef, no key sig, 4/4 time sig. Where has the key sig gone???
9. follows staff above
10. follows staff above
11. follows staff above
12. follows staff above
13. Indented staff, with a coda sign (should it follow the staff above?)
14. follows staff above

The staff 8 is marked as "SOLO". There are no notes, just chord names. Is this the reason for the lack of key sig?

[fxpar]

hi Hervé,

I hope you are fine, and sorry for the delay.

Key signatures
Yes that would absolutely make sense (sorry that it bothers you).
For me that part was not a problem at all, as my final conversion format is abc notation. The key signature being applied only once for the whole piece in abc I didn't jump at the error messages of audiveris.
But indeed, some books do not repeat the key and an explicit rule repeating it would be a nice addition to audiveris.

Solo
Indeed, some parts are meant to be done by only accompaniment chords. It could be used for background for improvisation.
(not sure if you are interested in this, but if you want some examples of comping rythms: https://www.fxparlant.net/variation-de-laccompagnement-en-abc-notation/ )

Thanks for your remarks.
Regards
François

[fxpar]

Hi, The new version (july 11) solves a lot of sharp and rest problems.
thanks a lot. 👍Great improvement. 😉

[hbitteur]

@fxpar
Thanks for your kind words.
The recent release improvement addresses the shape recognition.

I still have to work on the clef/key/time persistency from one staff to the other. Something more difficult.
Stay tuned!