Playscore 2 for Windows - Experiences

Dorico
1

I have tried most of the sheet music scanning programs out there over the years. None have ever been anywhere near satisfactory for my purposes. The best I have found was Smartscore 64. It consistently misreads notes on the top line of the staff (F for TC and A for BC). It also misses many other things such as multi-measure rests and chord symbols. But with practice I was able to get proficient enough with correcting the scanning errors within Smartscore, such that on some projects, it was a net time savings versus just entering everything from scratch. The support is basically non0-existent. If you do get a response from the proprietor, it might come across as “Well, obviously you aren’t smart enough to appreciate the program, so I’m done with you.”

In this world of AI, it seems like sheet music recognition is the perfect application because there is an infinite source of training data and the results should never be ambiguous. So, I was hoping that maybe one of the newer products would be a step forward.

I had seen Playscore 2, and I tried it on an Android pad, but I found the workflow completely opposite what I would want. The Android version is geared toward taking pictures with the camera and then transferring thugs to other platforms. This wasn’t a good workflow for me.

I recently noticed they have a Windows version that can process PDF files, which would be essentially the same workflow used by the other products. SO I decided to try it on a project – I actually have 10 big band arrangements I need to import into Dorico for further editing. Here are my observations:

THE GOOD

  • I didn’t have any technical problems. The Playscore app never aborted and always produced an XML file that was readable by Dorico
  • It was quite good at recognizing the notes, accidentals and such. I think in 30 scanned pages, there were only two notes missed altogether. Several others were interpreted as tied across a bar, causing the pitch at the downbeat to be off by a half step.
  • It did a good job with the articulations
  • It is fast. I didn’t do any editing inside Playscore, so I could process each part

THE BAD

  • I hate subscriptions. This is only available under subscription (about $8/month or $80/year). If I keep using it, I plan to turn off the subscription in the months when I don’t need it.
  • It doesn’t do multi-measure rests at all, and some other times, things music got shifted, but this is relatively quick to straighten out in Dorico.
  • It doesn’t recognize any chord symbols.
  • There is no editing capability within the program, and this is a problem because there is no good way to see any errors the program has made. Fortunately, it makes fewer errors than the other programs, and they were easy to fix within Dorico.

THE UNKNOWN

  • My source material had few slurs, dynamics or ornaments, which is part of the reason I am importing to Dorico, so I can’t say how well that works.
  • I worked from individual parts, not a score, so I have no idea how it handles scores or sheets with multiple instruments.
  • I had no complex repeat structures.
  • I had no meter changes, although once, Playscore cut a 4/4 measure short a beat.
  • My source material for this project was rather clean. I don’t know how it would do with rougher material.

Overall, I got better results faster than with the earlier scanning programs, but it is still far from a clean import. It is good enough for me to try a few more of my current projects. I found it encouraging, but don’t necessarily recommend it. Playscore does not claim to use AI. And indeed, if they are not using AI at the heart of the recognition process, they should be commended for not exaggerating about it. However, it it were using AI, I would be more encouraged because once an AI model is working at some acceptable level. it is often quick to improve the results by processing more representative training data. If it is not using AI, then it will probably remain at its current marginally useful level for years to come.

Here is a Scoring Notes article which covers several products.

2

I use PlayScore 2 quite happily on my iPhone.
I even like the fact I can prepare my pdf sources directly on the phone, even if I will do the importing of the resulting .xml file on the desktop computer.
Yes it is a subscription, but at a modest 28 Euros a year, which I am happy to pay for the quality of its results.
So my workflow is: open and look at the pdf on the iPhone, “share” to PlayScore 2, select the pages, let PlayScore 2 do its magic, then save as music.xml (to my cloud).
Then do the import to Dorico on the desktop at a later time.
Sometimes, if I am on Chaiselongue with iPad I can do both on the same machine, next to tea and biscuits…

3

I’m doing pretty much the same thing, it works great, but you’ll have to be/get comfortable with moving files around on your phone and between the phone and the computer. I use Onedrive for that, works great!

B.

4

The Windows subscription is nearly 3X that much. That’s looking borderline to me as far as value for the money. It does some things better than other programs, but it isn’t great.

I am unclear if AI is used at all or how much. This is important because if it is algorithmic programming, any further progress will be very slow going. But if it works this well using a neural net as the main engine, that is good news because NNs are “trainable”, whereas procedural code is not trainable. With procedural code, programmers have to come up with better algorithms, and that can be difficult with this kind of problem.

My guess is that Playscore 2 does NOT use NNs. If it did, we would probably see greater progress since the introduction. For example, if you look at the products that do “de-mixing” of audio files, most (or all) of them are NN-based, and they are making rapid improvements.

5

For whatever it is worth - AI is incredibly bad at skills like math - I guess you don’t need an insolvable philosophical question to defeat the evil AI overlords, just ask them to count triplets.

Anyway, the current focus they told us at my other job is less on teaching it skills but on teaching it to rely on external tools - sorta like telling a human to use a calculator or Google maps instead of teaching navigation. They seemed to favor an AI that could ask Playscore or Dorico to do the heavy lifting.

6

I think this is significant for all the OMR apps. Once you add the complexity of multiple staves in the system, things seem to go pear-shaped much more readily. Even piano music with polyphonic lines is often difficult.

Absolutely. Sadly, OMR isn’t where the money is…

1 Like
7

I never looked at it that way. I was going to use the Captain Kirk approach, but maybe I’ll just throw some triplets and grace notes at the next AI I encounter.