I don’t think this is part of Dorico’s core mission, nor should it be. Why would they go to the expense of retaining ‘experts’ to curate a potentially enormous collection of material?
What are you expecting the use to be? Pattern x is most often bowed ‘like this’? So what!
If you want to see the wide variety of bowings adopted by modern orchestras, just google videos of Berlioz’ March to the Scaffold!
Elgar, for one, used to write out particular passages many times with different bowings and play them through with his friend WH Reid until he found ‘the right’ articulation for his intent. But, hey, he was a violinist and understood what was possible.
Lastly, fashions change. Just look at the fingering and phrasing Tertis uses in his (authorised) arrangement of the Elgar Cello Concerto and compare it with what a modern Violist would do! Sorry, but we just don’t accept all that slipping and sliding around in modern performance practice.
Seems to me your AI engine would be severely compromised by ‘historical’ practices.
Exactly! I want the bowings in my music to reflect my style not some machine interpretation of average historical practices. It’s a formula for mediocrity.
No fingerings in music for professionals. String indications and use of harmonics etc are obviously fine. Pedagogical editions and Urtext editions will for different reasons not be suitable for AI involvement.
Bowings.
I will bow the same piece differently depending on things like: the predilections of the conductor; the performance speed; the acoustic properties of the venue; the number of players in the section; how capable the players are; and to be honest, my musical priorities, which evolve from year to year.
People who bow music regularly are used to overriding existing bowings and we will continue to do so if AI bowings start being supplied.
I don’t think that such a project will bear useful fruit (ie be an improvement on the current situation). But I will watch with interest to see if I’m proved wrong, though!
We all have to agree that such an algorithm can be same as good as a human being other than ourselves. Which means, we all do adjust someone else’s settings to our own tastes and we all can change our minds frequently.
What is the point of having AI add bowing between the source and the player, when the player will bow based on circumstances that cannot be known by the AI?
How would your AI exclude those people who produced a bowing specifically to reflect historic practice?
So what would be your criteria for that supervision? What precisely would you expect your ‘experts’ to do? Ignore historic practice? Compare historic practice? Reinforce current practice and reject historic practice? Create a new current practice?
That really depends on the ensemble. In a symphony orchestra, most maestros are insistent on uniform bowings that every player follows. It isn’t just for show – although that is a big part of it, but downbows have a different affect than up-bows, etc for all the different bowing techniques.
There are times that I have written three downbows in a row, which may be awkward for the players, but it forces them to think about really attacking the notes. For any music, the maestro might have very definite ideas about how many bows to use for a long note, when to hook and so on.
I am skeptical that an AI process could help much, but I wouldn’t discard it out of hand. I think this is the kind of exploration that is entirely appropriate at this stage of evolution of the notation products.
I’ve been battling a composer about fingering in their piano writing (some of which is actually my piano writing). We’re both conservatoire trained pianists, but we have different sized hands, differently shaped hands and different techniques. We also probably have a different idea of how the music should be voiced.
Regardless of the acoustic, the conductor, the instrument, the tempo, human beings are all different.
Can you explain how AI should account for the differences in us?
Yes, it is based on statistics. The algorithm needs the sheet music files to process in order to learn from them. It’s like an image recognition algorithm that gets trained to recognize a particular kind of rose. This is achieved by creating a folder and manually selecting let’s say 100 images of that particular kind of rose. This is how the AI gets trained and where the statistics come from.
This means, if you have average sized hands you get most advantage of it. This is like using monitors for audio engineering. We try to make it sound as equally good on as many devices as possible. This will never be 100% perfect. But it’s worth pursuing. I feel the same about fingerings and bowing of pieces I played which were provided by the publisher and we changed it to suit our individual needs.
As you use a piece of software your preferences can be taken into account. This is because your behavior constantly provides data to learn from and it is common to adjust the relative amount of data to be considered from different sources of data.
Why do you persist in arguing against your own logic?
You happily ‘changed’ the publishers’ suggestions to suit your own needs, but are arguing an AI solution will create a ‘right solution’ for all our needs??
@AaronVin you are the very definition of a hypocrite.
Sure - we’ve all told Google what a boat looks like in order to complete a captcha. It doesn’t know which rose or which boat I prefer, and despite that fact it knows what postcode I’m in it doesn’t seem to know that round here fire hydrants don’t stick up out of the ground.
But this, and the examples you’ve given, only take you as far as automatically recognising that a number between 0 and 5, without a rhythmic value to its left, is likely to be a fingering, which doesn’t take you terribly far at all.
Surely if you feed a bunch of different versions of the same piece into a database, what you get is an average of what was fed into the system, which may or may not be a representative average of the hand sizes of players today.
If three editions finger three adjacent notes (D-E-F#) as
2 4
2 4
-2-
and take an average of these, you’ll get the nonsensical
224, unless the system can guess that the first and second editors were probably going for 234 and the third editor was probably going for 123. Add into the mix situations where, even in scalic passages, editors sometimes do repeat the same finger (same goes for consecutive downbows or consecutive upbows) and you have something remarkably more complex than just recognising some blobs on the page and printing an average.
Again, you seem to be assuming that we’re all writing fingering for our own hands. Most of the time I’m typing in fingering, it’s for beginner violin music. I’ve never played the violin in my life - I’m just tidying up someone else’s work. My violin fingering probably applies to the vast majority of beginner violinists but wouldn’t apply to the bulk of professional players.
I’m failing to see how or why this should be within the Dorico development team’s remit. If you fancy doing a PhD on machine learning in regard to fingering - and it really does sound like a scintillating research topic - do go ahead! I’m sure Dorico will be a wonderful tool in your arsenal.
I don’t turn my music over to a person other than myself for bowing or fingering or what-not. I’m not sold that this is worth the kind of time it would take away from the Development Team’s work on far more useful features. You clearly think differently and are welcome to your opinion.
FWIW I’ve worked professionally with DNN recognition systems. This would be some kind of predictor. Instead of determining whether you have a ‘cat’ or ‘dog’ picture (a classifier’) in this case you’d recognize patterns and output a likely solution. It’s much like AlphaGo.
So the problem is data sets, you needs lots and lots and lots of data. More than you can afford, you need it not just for training but also for testing. Think hundreds of thousands to start with, millions would be better.
Why do we have predictors for images? Other than some fundamental technology (algorithm and hardware) it was the internet, and all the images getting uploaded. Voice and text is just as easy, and translation too. Google translate works just because we have so many web pages that are multi language.
Getting a network that works is as much an art as a science, but if you don’t have a cheap source of data forget it. According to OP ‘pay humans to mark up the scores’ - yeah exactly, I guarantee you won’t find budget for such an effort as the payoff won’t be great enough.