Transcript On Musical Stylometry—a Pattern Recognition Approach

Alex Stabile
Research Questions:
• Could a computer learn to distinguish between different
composers?
• Why does music by different composers even sound
different?
Possible Answers
• Backer et al.: On Musical Stylometry—a Pattern
Recognition Approach
• Analyzed low-level musical characteristics: note entropy,
intervals, rhythms
• Used information as input for a statistical model
Project Design
• Chords/harmonies all have their own character, so:
• Analyze harmonies found in music
• Use machine learning techniques to find a relationship
between types of harmonies and musical style
• Used Python, analyzed Midi files
• Compared works by Mozart to works by Rachmaninoff
Example File
http://www.ccarh.org/courses/253/files/midifiles-20080227-2up.pdf
Organization/Parsing file
•
Beat class
Notes on beat
(Beat number = 8)
Notes off beat
Chord Identification
•
•
•
•
•
Notes: C, E, G
What kind of chord? Look at intervals…
E: m3, m6
-no matches
G: P4, M6
-no matches
C: M3, P5
-These intervals form
a C major chord, root position
Analyzing Data—Machine Learning
Approach
•
Neural Networks:
•
•
•
Each node has a value and an
associated weight
Top layer is receives input
Values are propagated through
the network, creating values for
the other nodes
A simple neural network
Learning Algorithm
•
The network is given a set of training data whose outputs
are known
Inputs are “fed”
through the network:
Calculated output is
compared with desired
output to obtain error
http://galaxy.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html
Learning Algorithm
•
Back-propagation: the error is propagated
backward though the network, and a respective
error is calculated for each node
•
The weights and node values are adjusted based
on the errors so that a more desirable output will
be obtained
http://galaxy.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html
Learning Algorithm—This Project
•
Inputs to the network are the frequencies of
different kinds of chords
•
Two composers analyzed: Mozart and
Rachmaninoff
•
Expected output for Mozart: 0
•
Expected output for Rachmaninoff: 1
Results
4,000 Iterations
14,000 Iterations
10,000 Iterations
20,000 Iterations
Interpretation of Results
• Relationship between harmonic content and
style/composer
• Humans may learn to analyze this subconsciously, but a
computer can be trained to do so as well
Future Research
• Analyze more musical factors
• Analyze more composers
• Analyze composers who are more similar (e.g., Mozart and
Haydn)