ElmerJFudd Posted March 22, 2021 Share Posted March 22, 2021 In the 1980s... "Wabot-2 [was] a robot musician who could read a musical score by eye and perform the songs live on an electronic organ! This anthropomorphic robot was developed by a study group of the Waseda University's Science and Engineering Department to realize 'soft' functions of robots such as dexterity, speediness and intelligence. It was capable of playing with varying velocity on the keys, as well as foot pedals and was supposedly able to converse with people using its mouth (speaker) and ears (microphone)." [video:youtube] In 2021 we have AI that can drive a car (perhaps not perfected yet, but working in lots of situations). I'm surprised engineers haven't game-a-fied reading sheet music and taught an AI to perform and even interpret sheet music - by comparing notation to human performed recordings. And, assuming that is possible, why does Neuratron Photoscore Ultimate still suck so bad? You'd think by this time you'd be able to get dead accurate scans to MIDI or Notation file format. Anyone have experience with any of these? Have they made a jump in accuracy? https://www.playscore.co/blog/convert-sheet-music-to-midi/ https://scan-score.com/en/how-to-convert-sheet-music-into-midi-the-simple-way/ http://www.visiv.co.uk/ - these guys apparently were bought by Neuratron ------ Google DeepMind can play Break Out in 10mins. After 120 mins of playing it wins the game every time. [video:youtube] Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
GRollins Posted March 22, 2021 Share Posted March 22, 2021 I can't see why that should be al that difficult. Granted the AI's performance might be a little stilted, but you should be able to grant it a percentage of leeway in interpretation of note durations, etc. Hell, I've got a Eurorack sequencer that allows for note and timing variations--you assign percentage probabilities that it will vary and establish parameters as to how much variation is allowed. Should be a cinch. Make allowances for optical recognition of a handwritten score, though. Grey Quote I'm not interested in someone's ability to program. I'm interested in their ability to compose and play. Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 I can't see why that should be al that difficult. Granted the AI's performance might be a little stilted, but you should be able to grant it a percentage of leeway in interpretation of note durations, etc. Hell, I've got a Eurorack sequencer that allows for note and timing variations--you assign percentage probabilities that it will vary and establish parameters as to how much variation is allowed. Should be a cinch. Make allowances for optical recognition of a handwritten score, though. Grey I was thinking if you told it the goal was to sound like - piece played by professional musician. And if you kept feeding it more and more pieces and with the goal of sounding like a pro, it would eventually sound very much like a pro. Also, mechanical play of an instrument by a robot is cool, but not as important as reading/interpreting the score. Playback could just be from a synth engine with MIDI conversion happening after the fact. In the beginning, I'd say no handwritten scores. Give it sheet music input from PDF export of one notation app, like Sibelius or Finale. Then teach it to read output from various notation apps, then give it things written by hand. Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 Some examples of what current apps are doing. I don't own yet, so haven't given it something challenging. [video:youtube] [video:youtube] Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 Scan Score doesn't look any better - more accurate or easy to make corrections than PhotoScore Ultimate. [video:youtube] Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 This looks promising, but I have yet to find someone showing how frequently it makes errors. [video:youtube] Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
MathOfInsects Posted March 22, 2021 Share Posted March 22, 2021 The issue in traversing this final "uncanny valley" in score interpretation is that the human interpretation is only random or chaotic as compared to literal interpretation. That is, it's not really that we simply vary note-duration or tempo or dynamics, it's that we do it in service of some narrative or meaning we have created out of the elements in that passage or piece. If we just arbitrarily applied those elements to the pieces we played, it would be like hearing a nervous breakdown happen in real time. AI doesn't yet form a "theory of a case." It can only learn statistical likelihoods, and even if they're sometimes (coincidentally) situationally appropriate, they don't yet add up to a believable human performance. I feel like this and the day AI learns to craft its own joke from a situation, will happen roughly at the same time--which might be soon, but also might be never. But in the prefab-pop realm, machines are way ahead of us. If you're playing quarter-note bass pulses in loops, any variation is likely to be random. Might as well have the robot do it as the human. The algorithm might even come up with some cool things a person might not have gone for (though also vice versa, of course). Quote Now out! "Mind the Gap," a 24-song album of new material. www.joshweinstein.com Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 Yes, I understand. An AI performing sheet music and making artistic choices like a Glenn Gould or Charlie Parker with us debating which is more musical, the AI or the Master is unlikely today. However, we do see AI painting recreating a photograph as a painting in the style of X. Given enough definitions of the symbols we should be able to get a Sibelius like reading/playback of a score with minimal or no errors, no? Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
AROIOS Posted March 22, 2021 Share Posted March 22, 2021 Has anyone tried the Soloist (or Melodist) in Band-in-a-Box from over 20 years ago? Even in the late 90's, they already sounded better than 99% of the poor attempts at improvisation I saw (heard). And that's not even AI-based. It was just driven by a bunch of carefully studied patterns typical to a style/musician. On a related topic, I never cared for the improvisational part of (purist) Jazz. This is gonna offend a lot of people, but to me, the vast majority of Jazz improvisation sounded like public masturbation. The perpetrators surely enjoyed it, but it's not necessarily enjoyable for the audience, except those few who also enjoy musical-masturbation/mental-diarrhea. I'll take a well thought-out and rehearsed solo over 99.9% of improvisations on any given day. Quote Link to comment Share on other sites More sharing options...
MathOfInsects Posted March 22, 2021 Share Posted March 22, 2021 I'm honestly not sure. CAPTCHA, for example, works specifically by obscuring visual elements with enough noise that a data-scanner can't guess the underlying figures. It seems to me that would be the exact situation that would lead to interpretation errors from optical-scanned scores. For clean scores with clear markings and little noise, yes, I would think it's completely in the realm, even now. For noisy scores, or untraditional markings, or even the slightest bit of ambiguity around score-specific elements? I think the move to "error-free" would be MUCH harder to manage, even with mountains of machine-learning trial-and-error. I do have a friend I can ask about this. I teach him on Tuesdays. I'll ask tomorrow. Quote Now out! "Mind the Gap," a 24-song album of new material. www.joshweinstein.com Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 I'm honestly not sure. CAPTCHA, for example, works specifically by obscuring visual elements with enough noise that a data-scanner can't guess the underlying figures. It seems to me that would be the exact situation that would lead to interpretation errors from optical-scanned scores. For clean scores with clear markings and little noise, yes, I would think it's completely in the realm, even now. For noisy scores, or untraditional markings, or even the slightest bit of ambiguity around score-specific elements? I think the move to "error-free" would be MUCH harder to manage, even with mountains of machine-learning trial-and-error. I do have a friend I can ask about this. I teach him on Tuesdays. I'll ask tomorrow. Right. Digital noise from an imperfect score. Stuff that an experienced human reader could easily interpret based on what harmony is happening right now, or what the melody or bass line would typically do in the situation. An educated guess, and this is where current products fail miserably. Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
ElmerJFudd Posted March 22, 2021 Author Share Posted March 22, 2021 Has anyone tried the Soloist (or Melodist) in Band-in-a-Box from over 20 years ago? Even in the late 90's, they already sounded better than 99% of the poor attempts at improvisation I saw (heard). And that's not even AI-based. It was just driven by a bunch of carefully studied patterns typical to a style/musician. On a related topic, I never cared for the improvisational part of (purist) Jazz. This is gonna offend a lot of people, but to me, the vast majority of Jazz improvisation sounded like public masturbation. The perpetrators surely enjoyed it, but it's not necessarily enjoyable for the audience, except those few who also enjoy musical-masturbation/mental-diarrhea. I'll take a well thought-out and rehearsed solo over 99.9% of improvisations on any given day. I"ve had a few versions of Band in a Box over the years. Even then, just a massive amount of work must have been put into developing the styles - including generating solos. It would be a great library of data to start from when building an AI to read lead sheets in real time, that"s for sure. Quote Yamaha CP88, Casio PX-560 Link to comment Share on other sites More sharing options...
AROIOS Posted March 22, 2021 Share Posted March 22, 2021 Has anyone tried the Soloist (or Melodist) in Band-in-a-Box from over 20 years ago? Even in the late 90's, they already sounded better than 99% of the poor attempts at improvisation I saw (heard). And that's not even AI-based. It was just driven by a bunch of carefully studied patterns typical to a style/musician. On a related topic, I never cared for the improvisational part of (purist) Jazz. This is gonna offend a lot of people, but to me, the vast majority of Jazz improvisation sounded like public masturbation. The perpetrators surely enjoyed it, but it's not necessarily enjoyable for the audience, except those few who also enjoy musical-masturbation/mental-diarrhea. I'll take a well thought-out and rehearsed solo over 99.9% of improvisations on any given day. I"ve had a few versions of Band in a Box over the years. Even then, just a massive amount of work must have been put into developing the styles - including generating solos. It would be a great library of data to start from when building an AI to read lead sheets in real time, that"s for sure. Yes, PG_Music/Yamaha/Roland/Korg have all put a lot of work into analyzing and programming styles in their software/hardware products. In particular, the quality of sound, arrangement and mixing on Yamaha's pro arranger boards (Tyros/Genos/High-End PSRs etc) have surprised me on many occasions. And BIAB has made huge improvements with their audio-based styles in the last 15 years. Although arranger keyboard/software are not popular in the US, they are fantastic music practice and learning tools. Quote Link to comment Share on other sites More sharing options...
Bobadohshe Posted March 23, 2021 Share Posted March 23, 2021 It's all coming quick I'm sure. Quote Kawai C-60 Grand Piano : Hammond A-100 : Hammond SK2 : Yamaha CP4 : Yamaha Montage 7 : Moog Sub 37 My latest album: Funky organ, huge horn section https://bobbycressey.bandcamp.com/album/cali-native Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.