240 likes | 479 Views
Computer Generated Music. Amy Hoover COT 4810 04/19/08. Introduction. Computer-generated music sounds artificial Current systems battle between: Knowledge-based approaches Prescreen song for goodness based on “known musical rules” Sound trite, uninspired, unoriginal Other approaches
E N D
Computer Generated Music Amy Hoover COT 4810 04/19/08
Introduction • Computer-generated music sounds artificial • Current systems battle between: • Knowledge-based approaches • Prescreen song for goodness based on “known musical rules” • Sound trite, uninspired, unoriginal • Other approaches • Allow for more novelty • Sound messy, uncollected
Outline • Representing music • Deciding what is “good” • Problems
Representation • Representation: What is the best way to encode music? • How do we as programmers design a structure to represent music • Intuitive answer: Should music be represented by musical rules and encodings? • Less obvious • Functional relationships through Compositional Pattern Producing Networks (CPPNS)a
Representation: Mathematical Models • Probabilities • Idea: Actual musical notes distributed with certain probability, model with computer program • Without computer: Earliest music generation form (Mozart, Xenakis, Schoenberg) • With computer: Iliac Suite by Hiller and Isaacson (1947)
Representation: Mathematical Models • Probabilities • Idea: Actual musical notes distributed with certain probability, model with computer program Probabilities: A = .60 B = .40
Representation: Mathematical Models • Probabilities • Idea: Actual musical notes distributed with certain probability, model with computer program Probabilities: A = .60 B = .40 A B A A B
Representation: Markov Chains • Markov chain: “Conditional probability systems where the probability of future events depends on one or more past events” • Probability chart/state transition matrix • E.g.
Representation: Markov Chains • Probability of A following B (P(A|B)) = 0 • Probability of B following A (P(B|A)) = .5 A B B B B
Representation: Grammars • Idea: Music and language share similar origin • Relates sentence composition to music composition • Musical -> linguistic relations • Notes -> words • Phrases -> sentences • Melodies ->paragraphs
Representation: Grammars • E.g. – Grammar description • In: The interval between two notes (5th, 4th, etc) • Dn: Direction of interval • SEQn: Sequence • SIMn: Simultaneity • Example generative rule for this grammar • SIM1 -> SEQ1 +SEQ2 • SEQ1 -> (I5, D1)+ (I8, D1)+ (I11, D1) • SEQ2-> (I5, D2)+ (I8, D2)
Representation: Grammars SIM1 SEQ1 SEQ2 (I5, D1) (I8, D1)(I11, D1)(I5, D2) (I8, D2)
Representation: Neural Networks • Idea: Artificial neural networks abstractions of human brain, can “learn” music by example • Outputs • Pitch, timbre, duration • Network structure: • Recurrent • Compositional Pattern Producing Networks (CPPNs)
Representation: Neural Networks • E.g. • Network encodes note A on the Piano .21 .89 Pitch node -> <- Timbre map Pitch Map Example A <= .33 .33<B<=.67 .67<C<=1 Timbre Map Example Bass <= .33 .33<Guitar<=.67 .67<Piano<=1
Determining Goodness: Critic • Idea: Generated music sounds good, neutral, or bad • Critic: Agent to distinguish good and bad • Human • Rule-based • Learning • Evolved
Critic Types: Rule Based • Idea: Music mutated by good, musically sound rules • Given set of melodies, mutate according to musical rules such as • transposition, retrograde, inversion, augmentation • Typically brittle • Confined to style • No room for composers to break the rules • (Even dissonant chords can be resolved!)
Critic Types: Human • Individual/Group • Preserve novelty • Interactive Evolutionary Computation (IEC) • User presented with population • User chooses good individuals • Good individuals parent next generation
Critic Types: Example • GenJam • Generates jazz solos to “trade fours” with Biles • Jazz solo genomes are human evaluated as good or bad by a person • Winning genome accompanies Biles in real time
Critic Types: Human • Problems and Concerns • Truly creative? • User unreliability • User fatigue/ Human bottleneck • Many-to-one, obfuscates style
Critic Types: Learning-Based • Critics trained with “good” music • Learns to make “good” compositional decisions • Benefits: • A priori knowledge not necessary • Avoids human bottleneck • Disadvantages • Often over trains, does not generalize well
Open Problems in Evolutionary Music • Problem 1: • Current system designs are not recognized for artistic contributions • Problem 2: • Theories behind music/art systems are weak or non-existant
Conclusion • Computer generated music: Sounds artificial • Representation systems • Probability • Markov Chains • Formal grammars • Neural Networks • Critic types • Human • Ruled-based
Questions • 1. Name two musical representation designs • 2. What kind of critic does GenJam use?
References • Bently, Peter J., David W. Crone. Creative Evolutionary Systems. Academic Press, 2002. • Miranda, Eduardo Reck. Composing Music with Computers. Focal Press, 2001. • Miranda, Eduardo Reck, Al Biles. Evolutionary Computer Music. Springer-Verlag London, 2007. • Todd, Peter M., Gareth Loy. Music and Connectionism. MIT Press, 1991. • Todd, Peter M. and Gregory M. Werner. Frankensteinian Methods for Evolutionary Music Composition. MIT Press, 1998.