EMUSIC-L Digest Volume 42, Issue 05 This issue's topics: Neural Networks and Computer Music (10 messages) Your EMUSIC-L Digest moderator is Joe McMahon . You may subscribe to EMUSIC-L by sending mail to listserv@american.edu with the line "SUB EMUSIC-L your name" as the text. The EMUSIC-L archive is a service of SunSite (sunsite.unc.edu) at the University of North Carolina. ------------------------------------------------------------------------ Date: Tue, 14 Jul 1992 09:40:34 EDT From: "Joseph D. McMahon" Subject: Neural Networks and Computer Music (fwd) Peter Hesse writes: > A resounding good-day and hello to all of you out there in > the world of internet and bitnet. My name is Pete Hesse, > and I am a senior at the Thomas Jefferson High School for > Science and Technology in Alexandria, Virginia (Future home > of the Super Bowl Champion WASHINGTON REDSKINS!). I am > beginning a project with a fellow student, Aaron Berkson > dealing with the applications of neural nets in computer > music. We are working in a "Mentorship" environment, in > which Aaron and I travel to American University in > Washington D.C. throughout the summer and 2-3 days a week > during the school year. Our mentor, Eric Harnden is an > instructor of sound synthesis at American University. > He has suggested to us to use the full power of the internet > and pick the brains of all you that may be able to help us. > > My first question, or "favor" to ask is if anyone has > suggestions on good "readable" texts dealing with neural > networks. Also, does anyone know of publications, source > code, journal references, or FTP archives that discuss or have > information on previous projects dealing with the applications > of neural networks in computer music? > > I ask that responses be either e-mailed to > phesse@gmuvax2.gmu.edu > >OR< > posted in the newsgroup that you have read this as a reply. > > I thank ALL of you for your help and support, and look > forward to hearing from you! > > |) < - < > | < | < > Pete Hesse > SysOp of TechNet BBS > (703) 941-3572 > Welcome, Pete! Can you perhaps narrow down a little bit what you're after with this project? There may be some good stuff in back issues of the Computer Music Journal. --- Joe M. ------------------------------ Date: Tue, 14 Jul 1992 19:05:46 +0200 From: A.MULDER@ELSEVIER.NL Subject: Re: Neural Networks and Computer Music (fwd) Indeed there's quite a bit in CMJ's back issues on this topic. For a practical, yet not theoretically abandoned text on NN's try Russel Eberhart's book "PC neural network tools" (i'm not really sure the title is correct) published by Academic Press in 1991 or 1990. It includes programs even for music related stuff. But there's loads of books on NN's .... You can take your pick ! Then there's Neuron Digest and INNS-L (email lists) I get Neuron from mailbase@mailbase.ac.uk. This is not a standard listserver so take care what you say to it (try "help"). Neuron's quite good. INNS-L is not so active. Get it from Listserv@umdd.bitnet. Axel ------------------------------ Date: Wed, 15 Jul 1992 10:46:42 GMT From: Hans DU BUF Subject: Re: Neural Networks and Computer Music (fwd) In article <9207141340.AA00477@twinpeaks.gsfc.nasa.gov>, xrjdm@CALVIN.GSFC.NASA.GOV (Joseph D. McMahon) writes: |> Peter Hesse writes: |> > bla bla bla (sorry for the abbreviation) |> > |> > My first question, or "favor" to ask is if anyone has |> > suggestions on good "readable" texts dealing with neural |> > networks. Also, does anyone know of publications, source |> > code, journal references, or FTP archives that discuss or have |> > information on previous projects dealing with the applications |> > of neural networks in computer music? I saw a presentation at Neuro-Nimes 1991 in France, title "Neurswing: A connectionist workbench for the investigation of swing in jazz", about a system to change parameters in real-time. These were hot/cool, consonance/dissonance, and as-is-ness/substitutions, together with probabilities of harmonic, melodic, and rhythmic choices. Unfortunately, the presenter did not include a full paper in the proceedings, only an abstract. You should try writing him: Denis L. Baggi Xi Computer Corp. Breganzona Switzerland This address is far from complete, but will do here! Good luck, Hans du Buf ------------------------------ Date: Thu, 16 Jul 1992 08:33:48 CDT From: Phantom Tryke <"guinan::kummer"@KIRK.MSOE.EDU> Subject: RE: neural nets again >is there anyone on this list who has or is currently engaged in any projects >in the area of neural network applications in computer music? >if so, we have both specific and general questions, and would like to converse >with you, either directly or through the list, as seems appropriate. >second... can anyone provide specific journal or other references *other* >than the CMJ issues devoted to the topic. i have undertaken a preliminary >search through databases available to me, and have not come up with much. >on this question, it would probably be best to post directly to me. if the >response warrants, i will post a summary to the list. I, too, am _very_ interested in applying neural networks in computer music. There is one book that I know is available that I am going to purchase as soon as I get the $70: Neural Network PC Tools. by Eberhart. Sorry I don't have the ISBN number or publisher info. As I recall, this book is practically dedicated to computer music and is loaded with C source code for some sample programs. phantom tryke kummer@kirk.msoe.edu ------------------------------ Date: Thu, 16 Jul 1992 08:43:36 EDT From: ronin Subject: neural nets again ok, let me make this a little crisper... is there anyone on this list who has or is currently engaged in any projects in the area of neural network applications in computer music? if so, we have both specific and general questions, and would like to converse with you, either directly or through the list, as seems appropriate. second... can anyone provide specific journal or other references *other* than the CMJ issues devoted to the topic. i have undertaken a preliminary search through databases available to me, and have not come up with much. on this question, it would probably be best to post directly to me. if the response warrants, i will post a summary to the list. -----------< Cognitive Dissonance is a 20th Century Art Form >----------- Eric Harnden (Ronin) or The American University Physics Dept. 4400 Mass. Ave. NW, Washington, DC, 20016-8058 (202) 885-2748 ---------------------< Join the Cognitive Dissidents >------------------- ------------------------------ Date: Thu, 16 Jul 1992 17:45:55 -0600 From: Thunder-Thumbs Subject: Re: neural nets again --------Here it is-------- >ok, let me make this a little crisper... >is there anyone on this list who has or is currently engaged in any projects >in the area of neural network applications in computer music? >if so, we have both specific and general questions, and would like to converse >with you, either directly or through the list, as seems appropriate. -------That was it-------- A computer science professor here (CU-Boulder) is doing such a project, and I'm helping him. However, I don't really know a lot about it yet, because it's a new project, I just started, and I am currently better at the music end than the computer end. I have some documentation, but I don't know if he wants it released or if it's available by ftp. I can find more out if you are interested. Anyone interested should write me back at siffert@ucsu.colorado.edu. Thanks, Curt Siffert ------------------------------ Date: Wed, 29 Jul 1992 11:01:30 EDT From: ronin Subject: nets continued ok, this is my last one of the day. i decided not to combine these three posts, since they really are on pretty different subjects. you may remember my asking (and my students' asking) recently about neural nets. we did get some responses, for which we are thankful. clearly, we're going to have to buy those two books, Music and Connectionism, and The Well Tempered Object... they seem to be the going texts. but before i do buy them, let me just lay a little question out, and see if i can get some naive brainstorming. no references, please... i can find them myself. i'm interested in what the creative people in this group think about this... one of the critical issues in neural network research is that of data representation. simple kinds of information, such as numeric data of a given distribution, might be able to be presented to the network 'raw', but the more complex the information from which one wants to extract features (or classify, or whatever one is trying to with it), the more pre-processing is required. after all, most nets consist of but a tiny fraction of the many layers of highly interconnected units that make up our sensorium, processing, memory, and other mental systems. lacking these, a comparitively simple artificial neural network needs to have its input data preprocessed, perhaps pre- or re-organized for it, as a given cortical region might take its information not directly from perception, but from prior neural areas. an example of this lies in image processing, in which one scans the image in some way more closely resembling the way the eye shifts its attention around features of interest in a scene, rather than simply handing a processing network a raw raster scan. the question, then, is what you might come up with as a way of representing music. again, i know that this question has been raised, researched, and published on. and i will look up the results. but i want to know what might occur to you, not to the formal research community. so what would you do? how would you represent the patterns that are so clear to you as a musician or an audience? when looking at a score, your eye does not see statistical properties. it registers motion, density, uniformity. midi data does not, to the reader or to the programmer, communicate music except by several layers of internal translation. and is there a rough isomorphism between how the eyes and ears interpret the score and the music? to my mind there is. if i were to describe a music without reference to style, i would think in terms of motion, density, and uniformity, as well as consonance, which does not seem to me to be so clear on the page. would i then represent notes in some manner that preserves something of their relations (since it the relations among things that defines pattern)? specify an event, not in terms of its absolute pitch and duration, but in terms of its difference with respect to the last event? can one encode a macro-structure, a tendency? and so on. what do you think? -----------< Cognitive Dissonance is a 20th Century Art Form >----------- Eric Harnden (Ronin) or The American University Physics Dept. 4400 Mass. Ave. NW, Washington, DC, 20016-8058 (202) 885-2748 ---------------------< Join the Cognitive Dissidents >------------------- ------------------------------ Date: Wed, 29 Jul 1992 11:42:34 EDT From: metlay Subject: Re: nets continued > so what would you do? how would you represent the patterns that are so clear > to you as a musician or an audience? when looking at a score, your eye does > not see statistical properties. it registers motion, density, uniformity. Hmmm. You're making an eye-ear-brain connection here. I don't know if that extra link in the chain is either necessary or helpful-- after all, when *I* look at a score, I see a bunch of funny lines and dots. I see motion and density only in the very vaguest sense, and have no cognitive link between ink-blots and notes. > midi data does not, to the reader or to the programmer, communicate music > except by several layers of internal translation. and is there a rough > isomorphism between how the eyes and ears interpret the score and the music? Some may argue that there is-- that is why people love to interpret music scores that are drawn in a non-traditional manner, because it short-circuits the connection to learned behavior patterns and puts the imagination in its place. But a computer has no imagination, yet literal translation isn't what you're after either.... > to my mind there is. if i were to describe a music without reference to style, > i would think in terms of motion, density, and uniformity, as well as > consonance, which does not seem to me to be so clear on the page. would i then > represent notes in some manner that preserves something of their relations > (since it the relations among things that defines pattern)? specify an event, > not in terms of its absolute pitch and duration, but in terms of its difference > with respect to the last event? can one encode a macro-structure, a tendency? > and so on. Depends on the music style. Abrupt transitions aren't modeled well by Markov chains or the like. I think you'd have an easier time with trane music than with PDQ Bach.... > what do you think? I think it's a more interesting topic for discussion than which workstation synth to buy next, frankly. |-\ -- mike metlay (phud, no less) metlay@minerva.phyast.pitt.edu ------------------------------ Date: Wed, 29 Jul 1992 13:14:45 EDT From: "Joseph D. McMahon" Subject: Re: nets continued ronin writes: > so what would you do? how would you represent the patterns that are so clear > to you as a musician or an audience? when looking at a score, your eye does > not see statistical properties. it registers motion, density, uniformity. > midi data does not, to the reader or to the programmer, communicate music > except by several layers of internal translation. and is there a rough > isomorphism between how the eyes and ears interpret the score and the music? > to my mind there is. if i were to describe a music without reference to style, > i would think in terms of motion, density, and uniformity, as well as > consonance, which does not seem to me to be so clear on the page. would i then > represent notes in some manner that preserves something of their relations > (since it the relations among things that defines pattern)? specify an event, > not in terms of its absolute pitch and duration, but in terms of its > difference with respect to the last event? can one encode a macro-structure, > a tendency? and so on. I remember a remark by Doug Hofstader about Chopin - after listening to the music a lot and deciding he wanted to try to play it, he bought the scores to the preludes, only to find out that he was facing what he called "grinning skeletons" of the music. If you're looking into the design of a score-to- MIDIfile translator, then scores are worthy of analysis, but if you want a music listening-program, you are being seduced by the the power of vision over our other senses - I'm not sure that it adds anything but complexity. Now, feeding the program a MIDI file to "read" while it listens may be more useful, since there's a lot of information there which *has* been processed for you - the pressure and velocity data, precise note data, and other information which makes it easier to "get inside" the performance. Maybe this is too much information...? However, speaking as a improviser and musician, this doesn't have a lot to do with how I listen to a piece of music. The primary thing that I hear in music is patterns and relations. Even in the most abstract music, a recurring rhythm or musical figure can be latched onto to help establish a context in which the piece can be rooted. It seems to me that the analysis of a piece of music is sort of like watching a movie: the actions of the players (pun intended) are defined in relation not only to the background, but to each other. An action in one context may have a completely different implied meaning in another context. A music listener needs to be able to shift focus from "foreground" to "background", and to recognize things in obscure contexts. Try listening to a Bach fugue. You can either follow a single voice, in which case the others "blur" into a harmonizing background, or you can follow the overall flow all at once; but it is intensely difficult to listen to two voices and to try to follow them simultaneously. Perhaps you will want multiple listeners, who can compare notes? Stockhausen says that Kontakte (the all-electronic version, at least) is based on multiple scales, which use notes which cover wider and wider spectral ranges, and on rhythms which are shifted from very long periods to audio frequencies. I caught the second part and had verbalized it to myself, but until I read his comments on the piece, I did not realize the first part. Going back and listening again with this information made it much more obvious what was going on in the piece and why it is as long as it is. How does one represent such meta-information about the piece? Certainly gaining it greatly increased my understanding of how the piece worked and why. The question here, I guess, is how would one deduce such a thing? I certainly missed them, and I think I'm a relatively sophisticated listener... --- Joe M. ------------------------------ Date: Thu, 30 Jul 1992 09:08:16 +0100 From: Martin Roth Subject: Re: nets continued Eric Harden (Ronin) writes: > how would you represent the patterns that are so clear > to you as a musician or an audience? when looking at a score, your eye does > not see statistical properties. it registers motion, density, uniformity. You are talking of "patterns", as dealt with in any pattern recognition program (PR). This means you can not just run a statistical analysis of the raw data and come up with a measure of 'motion'. You will have to group some events while excluding others to see the motion (as I understand, there may be several 'motions' in a piece at once, say a slow chord progression and a part of a scale as melody - if _that_ is not 'motion', sorry, I don't know much about music theory, and I'm not native english speaking). Currently, I'm writing a PR (pattern recognition) program as my diploma thesis which is able to 'read' printed music (sheet music, scanned in). This seems relatively simple since there are lot of rules on notation. But then, as you put it, 'your eye' sees that there are lines, heads, rests. You recognise the symbols 1) without any 'thinking' ans 2) even if a symbol is partly occluded or touches others. Now programs do not have this ability, and there is no 'standard' way to deal with this. Finding patterns in music is probably even worse than finding 'optical' patterns. It's difficult to find a measure for grouping events (in optical PR, neighbouring pixels are grouped). The problem is, you 'just see' (or hear) figures, ornaments. You can name some (chords, scales, intervals, pedal tone, melody). But you can't tell _why_ you think that some tones belong to the melody. In some peaces, the melody moves from one instrument (voice) to another, maybe even in the middle of a phrase. I would begin with the simplest task you can think of, for example recognising the chord (key?) of the music at every moment. This does not seem trivial to me if you run the program on a complete score. Of course it's easy the way keyboards with automatic bass, background and rhythm do this. But their chord recognition sees only the left hand notes and not the whole melody. Well, just my thoughts. I may be understanding 'motion', 'density' and 'uniformity' completely wrong... What do you want to do with the program anyway? 'Interpret' music? -Martin _______________________________________________________________________ _ Martin Roth Martin Roth ETHZ, ips, RZ F16 |\ /|_) Mail: roth@ips.id.ethz.ch Sandacker 14 g 01/256 55 68 | \/ | \ (Student of Computer CH-8154 Oberglatt p 01/850 32 75 Science / Engineering) Switzerland (F-)emails welcome! ----------------------------------------------------------------------- ------------------------------ End of the EMUSIC-L Digest ******************************