Musical ExpressionNo Longer Exclusively HumanBy:Danielle FrantzPapersA Survey of Computer Systems for Expressive Music PerformanceNeural Network Based Systems for Computer-Aided Musical Composition:Supervised x Unsupervised LearningYQX Plays ChopinOutlineIntroductionA survey of Computer Systems for Expressive Music Performance.Computer Aided Musical CompositionYQX Plays PianoEthicsMusical TermsTempo - The speed at which a song is playedMeasure - A segment of a song separated by bar lines. Dynamics - Represent the volume at which the notes are played.Articulation - Refers to how the notes are played (short, long, smooth, etc.)Vibrato - Fluctuation in pitch in order to produce a more pleasing sound.Introduction to Systems for Musical Expression1980's - MidiRobotic, Metronomic SoundsWhy do they need expression?Instead of making it humanlike, we turned robotic.(Synth-pop) Robotic music was RARELY used in classical music.Expressive Performance ActionsRefers to the actions of musicians NOT in the score.Tempo and Loudness (NOT in score)Expressive Performance ActionsExamples:Vibrato, Intonation, Slowing at the end of a piece, articulationWHY?Embellishment, InternationalismPerformance ContextExpression of a certain mood or emotionHappiness, Sadness, etc.Change tempo and dynamics to change the emotion.CSEMPComputer Systems for Expressive Music PerformanceAutomated - Once Trained, it generates a new performanceSemi-automated - Requires manual input. (Musicological Analysis)Computer Expressive PerformanceWhy would we want a computer to give an expressive performance?Develop computational models (Musicology, Psychology) Composing tool Expressive Computer Generated Music (Generate mood music for a computer game)Playing data files (MIDI)Computer Accompaniment (Track and Generate Expression)OutlineIntroductionA survey of Computer Systems for Expressive Music Performance.Computer Aided Musical CompositionYQX Plays PianoEthicsFramework for Research in CSEMPModulesPerformance Knowledge (Rules controlling the system)Music/Analysis (Analysis of structure: Automated, Manual, or a Combination)Performance Context (Emotion or Style)Adaptation Process (Develop performance Knowledge)Performance Examples (Past performances)Instrument model (Post commonly the piano)Methods of CSEMPNonlearning Linear RegressionArtificial Neural NetworksRule/Case-Based LearningStatistical Graphical ModelsOther Regression MethodsEvolutionary ComputationNon-LearningDirector Musices (Project since 1982)Phrase Arch (Rainbow Shape)Groups Tempo and DynamicsSound trails off toward the end of a groupWorks in Real Time Hierarchical Parabola ModelRelates Tempo and Dynamics similarly to DMThe Louder the FasterNot ConvincingBach Fugue SystemConditions Generated from an annotated knowledge base Works for polyphony (polyphony is a texture consisting of two or more independent melodic voices)Commercial ProjectsSibeliusProgram for music compositionBuilt-in algorithms for expressive performanceRules and TechniquesHigher the louderUse bar lines and beam groupings to determine emphasisRandom fluctuations added (More human-like)Notion and FinaleMore advancedFewer details about methodologyLearning MethodsCan incorporate more knowledge more quicklyNot as much creativityFocus on past performancesIn order for a CSEMP to learn to play expressively, it must be trained with both robotic and expressive performances in order to learn the deviations.Linear RegressionAssume linear relationship between music features and expressive actionsAdvantage:SimplicityDisadvantage:Almost certainly oversimplifiedMusic Interpretation System (MIS)Generates expressive performances in MIDI formatLearns expressive rules from audio recordingsLinks between music features and Performance actionsTempo, Dynamics, ArticulationMIS was trained on the first half of a Chopin waltz and generated an expressive performance for the second half.Neural NetworksInspired by the structure and processing of the human brainThe Network contains neurons connected together by synapses.Each connection has a weight (Referred to as the interconnection weight)Training PhaseFeed the network examples Testing PhaseGenerate guesses based on the training SetSupervised Vs. Unsupervised LearningSupervised LearningInput-Output Mapping Unsupervised LearningOnly Inputs KnownNeural NetworksStart with random weights. Goal is for the weights to converge to a line separating Red from Green.OutlineIntroductionA survey of Computer Systems for Expressive Music Performance.Computer Aided Musical CompositionYQX Plays PianoEthicsIntroductionNeural-Network Based Systems for Computer-Aided Musical CompositionCompares Supervised Vs. Unsupervised LearningUses 'Nature' as inspirationLandscape Contours are converted into integers Focus on Pitch and DurationNature representationMeasure represents 4 time units.Eigth Note: Quarter Note:Half Note:Whole Note:Integer RepresentationRepresentation of PitchCombine integers and intervalsIntegersFor Example: Middle C is 0 Each half step above middle C gets a new positive integerEach half step below middle C gets a new negative integerIntervalsDistance between two notes+ or - half stepsPitch and DurationSupervised LearningTraining Samples: Brazilian Folk MusicTrained with measuresOnce trained, compose new melodiesSupervised Using Back-PropogationThe network output is compared to the desired output. The difference of the two represents the error, which is then used to adjust the weights.Repeated until some stopping criteria... Example. Threshold or some number of iterations.Unsupervised LearningUnsupervised using a self organizing map (SOM)Cannot calculate error in the training stage.In each iterations, a winner neuron is chosen to update its weights.Weights are updated by adding the Learning rate times the difference between the input and the winner.The neighbors of the winner neuron are also updated by using the Euclidean distance between the winner and its neighbors.OutlineIntroductionA survey of Computer Systems for Expressive Music Performance.Computer Aided Musical CompositionYQX Plays PianoEthicsOverviewContest to create a system to learn to play piano expressivelyComputer has to play two pieces 'expressively' that it has never seen before The two pieces were composed specifically for the contest (in the style of Chopin)Scores contain normal markings, but no indication of phrasingContestants have 1 hour to set up their computer, read the scores and generate a
View Full Document
Unlocking...