Forget Scan2Notes. The answer is SoundSlice!

Dorico
1

I have been searching for a decent music scanning app for decades. There are lots of half-arsed attempts out there, but none that were anywhere close to commercial grade. Here is a note I posted earlier.

See Scan2Notes - The Most Accurate AI Sheet Music Scanner | klang.io

It is an online service that purports to scan music accurately – they claim the best accuracy ever.

The problem is it is subscription only and the subscriptions make no sense – limiting you to 50 scans per month. They have no free trial. They have a demo mode, but that only does 12 measures, and those 12 measures are pretty slow (several minutes). I did a test, and it did well on the 12 measures. But with a limit of 50 scans a month, I’d use that up in a couple of days sometimes. I have been incrementally converting a big band library, and my workflow is to scan each instrument as a separate job, then bring them into Dorico, one instrument at a time. I might only get 3 charts converted with a limit of 50 scans per month.

Beyond that, I can’t tell anything from a test of 12 measures.

Has anybody gone down the path with this service?

After I posted that, I stumbled upon Soundslice. Although I did not recall, my notes show that I tried it 2 years ago and I had set up a test account. I guess I wasn’t that impressed in 2024. But I tried it again, and it is looking great – so much so that I have started a subscription account to give it a good look.

I really, really don’t like subscriptions, but this one might be good enough to put up with the hassle of a subscription. SO far, I have only run a few parts through, but already, this is 100X better than anything else I have seen. The note recognition is perfect except on one measure where my scaller sessed up a couple of bars badly. Unlike any other program, so far Soundslice is cating all teh multimeasure rests properly, and also handiling 1st and 2ns endings. It does not process segnos, d.d., Coda, but those are easy to add in Dorico. Other than that botched scanner thing, my measure counts are exactly right. And it is even making good guesses for tenotos and staccatos that were written in by hand.

I scanned a bass part that had all the chord changed. After the initial processing, it pops up questions about any of the objects it was unsure about. It asked me if an A7#11 was correct. I replied that was correct, but strangely, it left those out of the MusicXML even though it correctly recognized them.

If this holds up, this will greatly accelerate my transcription work. I should be clear that I am scanning parts only. I have no idea how well it might handle a score or parts with multiple voices, so it may not work so well for everybody.

But this is really encouraging. The few dynamics that were in the original music came through. All the accidentals. All the double-bars.

This one has a subscription that is a little more generous than Scan2Notes. With Soundslice, you can process 100 pages of music per month on the “plus” plan for $5/month (or $50/year.)

2

I’ll agree that the Scan2Notes subscriptions make no sense. And “per month, billed annually” requires clarification.

3

I’ve just tried it on two pages, and I have to agree. This solution needs to be taken into account nowadays. Especially since I’m not sure for how long I will be able to use Photoscore Ultimate (last version is 6 years old), and MacOS sends me a message saying it won’t work on new versions of the system…

[But I don’t like the subscription model]

4

This has become very standard on websites. It’s really deceptive. You can only pay the full annual amount but they quote the amount divided by 12 as the monthly fee. Well yes its the monthly amount, but you cant pay monthly. And most of these places no longer offer a month by month plan. I really dislike it, and it often prevents me from taking up a product.

5

Hi, Adrian from Soundslice here. Glad you found our scanning feature and like it!

We are continually improving it and adding support for detecting more notations. And I assume this won’t necessarily be relevant to hard-core Dorico users, but: we do have a full notation editor built in, along with excellent practice tools. (The main point of Soundslice is to be a learning/teaching tool, and we developed the scanning feature to help people get music into their account more easily. It’s turned out that lots of people use our scanning feature independently due to its generally good results.)

Cheers from Amsterdam -

Adrian

6

Just gave it a try. Meh, not impressed. It prompted me for almost all the chord symbols, so at least those are correct, but not any faster than just inputting them myself in Dorico. It didn’t catch that this was a grand staff so it’s kind of a mess:

soundslice
soundslice1920×1298 340 KB

This was a pretty simple example and fixing it would definitely take longer than simply inputting it correctly the first time. I don’t think I can export on the free version so not sure how the XML would have imported into Dorico.

7

I didn’t try anything with a grand staff, any scores, anything with divisi or ossias. I expect those would be a mess.

On the pages I have done so far, the XML comes through well, with what you see in Soundslice essentially matching what ends up in Dorico.

The reason I am optimistic is twofold:

  1. My needs are simple enough that it is already faster than re-entering manually (although I suspect you are 3 or 4 times faster entering into Dorico than I am. :slight_smile:

  2. This is a proper AI system with what I believe to have legitimate machine learning from the users of the product. That is to say, when the program is unclear about things, it asks the user to clarify the correct interpretation. If these answers are folded back into the neural network, then the engine should improve as customers use the product, much like what we see with tools like Grammarly.

8

Good to meet you Adrian, and I am happy to see the progress you have made so far. I think you are correct that few Dorico users would be interested in using the editor. With that said, your subscription program doesn’t work very well for my circumstances. I am in the process of re-notating a book of 300+ big band arrangements. Some were handwritten and I simply start over with those, updating and modernizing the arrangements as I go.

Most were written in Finale, but unfortunately the originals were lost with a computer failure with no back-ups (not my computer – I have 27 layers of back-up, having been an IT professional from the mainframe days.) Some months I have other priorities and do none of this. Other months I might convert 5 or 6 charts. A typical chart (not including score) will have 17 parts, and they will all be 2 or 3 pages each. So one song could be 40 pages to scan. I might be able to do 3 smaller ones before you cut me off. But then I might go several months with nothing, so it makes no sense to pay the “teacher” price to get 200 pages. And there could be a month when I’d want to scan more than 200 pages.

I suggest you consider a perpetual license where the engine could be downloaded. That would enable you to sell upgraded versions periodically as the engine improves.

If you don’t like that, at least consider a credit-based system that is not time-based. Let me buy 100 pages of conversions in advance. When I use that up, I’ll buy another batch of credits, but you don’t hit my bank account every month or year. I only pay for what I use.

And if you want to limit that plan to just document conversion and review, that’s all I would be doing anyway.

9

If true, I’m optimistic about this. Other machine learning AI for music, like stem splitters, are amazing and have completely changed my transcription workflow.

10

It’s generally a discounted price when you pay for a year up front. But if there’s no monthly payment option, then it doesn’t make much sense.

11

It’s a tactic to become profitable very quickly. I suspect many developers have short-term loans that have payments that become due very soon after they release their software.

12

Sorry about that! You happened to stumble upon an edge case there. :neutral_face:

The issue was that your original music didn’t have clefs in the subsequent systems (after the first system), nor was there a systemic barline in the first system. That was a bit of a perfect storm.

The system does indeed work with multi-staff music. But at the moment, we require those two things (a clef in each staff and a clear systemic barline connecting them) in multi-staff music, to know how to interpret the multiple staves. Generally this isn’t a problem for 99% of people, but you got (un)lucky.

I wouldn’t call it a simple example, BTW. Speaking as somebody who’s built a music scanning system, “simple” in this context means “all the standard notations are in the expected places, and nothing is omitted.” Your case was definitely not simple.

Adrian

13

Thanks for your reply! When I was selecting a “simple” example to test, I meant simple as in very few notes, and something I could input correctly into Dorico myself in about 5 minutes. I was looking for a handwritten example just to see how it would do. I knew the first “New Real Book” from Sher was hand copied so just picked something from it.

As it’s a very popular book, its house style for those elements you mentioned are pretty commonly found. I personally am not a fan of not restating clefs and key sigs, and I agree the lack of an initial systemic barline looks odd, but the book is pretty consistent in presentation, even if it’s not my personal preference:

beautifullove
beautifullove1832×1488 275 KB

14

If you want some additional hand-copied training data, I think this entire collection is of mid-century Victor Young conductor’s parts is excellent:

Usually one of the most senior copyists on a team did the conductor’s parts as those paid the best and required the most attention. I just tried a random page from “Three Coins in a Fountain” and … it did not fare well. I’m out of trials for the month, but these aren’t just a composer’s scribble, but some of the best hand work you’ll find. It would be great if stuff like this could be accurately interpreted!

15

I have been using the legacy Photoscore for years, with better results than I have been able to achieve with its competitors’ trial versions–up until recently. Having not searched for something better for a year or so, but having noticed that AI intelligence is improving exponentially, I tried Playscore 2 recently. Although I have canceled participation in programs such as Adobe apps, Sketchup, and Sibelius when they switched to subscription only (yes, Sibelius has rethought that disastrous mistake), I found that Playscore 2 is far better than anything else I’ve tried and promises to become even more accurate in the fullness of time. I would say that it takes 10% of the time correcting the remaining mis-identified elements that would result from Photoscore OMR. One particularly difficult mistake to correct with Photoscore OMR is the incorrect assignment of staves that do not appear in introductory passages. Photoscore assumes, almost always incorrectly, that the first staves to appear should be assigned to the same instrument (s) as top staves that appear later in the score. The result is that scores have to be broken into pieces and instruments reassigned whenever a new instrument appears. In contrast, for instance, Playscore 2 figures out that the piano intro should be assigned the sames staff as the piano appearing below new instruments appearing later. Also, the OMR occurs on my machine locally (like Cantai, yay!), so the recognition latency doesn’t depend on the bandwidth of an engine in the cloud, as is the case with many AI driven apps. Overall it has a much higher accuracy in converting many musical elements in scores than anything else I’ve tried, and I am happy to deal with the subscription model as long as they provide this much improvement in the time it takes to fully capture a score. At this moment, I would encourage anyone looking for better OMR to try out the Playscore 2 demo. I have no dog in this fight, but I really like spending a lot less time correcting OMR mistakes.

16

You can reassign the staff numbers/instruments before reading. (Always the first task to check!)

(Though I’m not defending PhotoScore - it’s definitely showing its age these days)

17

I agree with @Janus . It looks like you have not learned properly how to use Photoscore Ultimate — and it’s not surprising, as learning it properly does require a real will from the user to do so… Talk about intuitive :joy::joy::joy:

My experience with Playscore2 is probably the same as yours, and we share the same love for subscription…

I’ve tried that other software that was mentioned earlier (Audiveris), open source. It looks it needs to be learned, just like Photoscore. I’ve tried it from scratch and was just able to waste some time, so I downloaded the manual (too long to read, but I will try to anyway).

18

No one on this thread has mentioned SmartScore. Anyone own or tried out the demo?

19

SmartScore’s main advantage is that it comes with its own editor, where you can correct errors before you export as MusicXML – and in fact it highlights bars where there are too many notes, etc.

However, it still gets too many things wrong to start with; and it’s still an Intel-only app on Mac, suggesting that development is not at a prodigious pace.

When I’ve contacted the company, they are always keen to point the cause of any problems as the images/scores that I was scanning, rather than their software.

20

I use Dorico to create scores based on sets of parts which generally look like this:

image
image1698×856 216 KB

I have just assumed no AI tool will be of use at this point based on my attempts at having ChatGPT 5.4 analyze it (and it completely failed even for a lot clearer examples). But perhaps I was wrong. Do you guys who have already sank time into explorations of new software think it might be worth to explore some software or am I better off doing everything manually? Thanks in advance!