240 likes | 437 Views
jSymbolic. Jordan Smith – MUMT 611 – 6 March 2008. Overview. jSymbolic extracts high-level features from symbolic (MIDI) data. Walkthrough of the interface Features: Types of features Motivation for choice of features Extraction Planned improvements. Overview.
E N D
jSymbolic Jordan Smith – MUMT 611 – 6 March 2008
Overview jSymbolic extracts high-level features from symbolic (MIDI) data. • Walkthrough of the interface • Features: • Types of features • Motivation for choice of features • Extraction • Planned improvements
Overview jSymbolic extracts high-level features from symbolic (MIDI) data. • Walkthrough of the interface • Features: • Types of features • Motivation for choice of features • Extraction • Planned improvements
Features • 3 kinds of features: • Low-level • High-level • Cultural
Features • 7 categories of high-level features: • Instrumentation (20) • Texture (20) • Rhythm (35) • Dynamics (4) • Pitch statistics (26) • Melody (20) • Chords (28)
Features • Why so many features? • Ensure ability to discriminate as many different kinds of music as possible • Want features to be as basic as possible, because: • They are destined for a machine learning experiment • Estimating complex features is controversial
Features • Why pick these features? • Long history of musicological interest • Relative ease of extraction
Features • Why pick these features? “The features described above have been designed according to those used in musicological studies, but there is no theoretical support for their … characterization capability.” (Ponce de León. 2004. Statistical Description Models for Melody Analysis and Characterization. ICMC Proceedings 149-56.)
McKay & Fujinaga 2005: Automatic music classification and the importance of instrument identification. Proceedings of the Conference on Interdisciplinary Musicology.
Overview jSymbolic extracts high-level features from symbolic (MIDI) data. • Walkthrough of the interface • Features: • Types of features • Motivation for choice of features • Extraction • Planned improvements
Using the Features • Like jAudio, modular features make it easy to add new ones -- ADDING FEATURES -- Implement a class for the new feature in the jAudioFeatureExtractor/MIDIFeatures directory. It must extend the MIDIFeatureExtractor abstract class. Add a reference to the new class to the populateFeatureExtractors method in the SymbolicFeatureSelectorPanel class. • Features exported to ACE XML or Weka ARFF
Feature Extraction • Other than jSymbolic, what is the state of the art in symbolic feature extraction? • Borrow from others or invent your own, and implement them by yourself. • Use MIDI Toolbox.
MIDI Toolbox vs. jSymbolic • jSymbolic -requires JAVA -is strictly for extracting features -analytical goals: usefully and objectively condense information • Toolbox -requires MATLAB -has tools for manipulating and visualizing data -analytical goals: estimate a musicologically important feature
Planned Improvements • Boost number of features from 111 to 160 • Ability to operate on non-MIDI symbolic data (MusicXML, GUIDO, kern) • Ability to extract over windows
References jSymbolic overview: • McKay, C., and I. Fujinaga. 2006. jSymbolic: A feature extractor for MIDI files. Proceedings of the International Computer Music Conference. 302-5. Details of features implemented in jSymbolic: • McKay, C. 2004. Automatic genre classification of MIDI recordings. (M.A. Thesis, McGill University). Example of jSymbolic’s feature extraction in action: • McKay, C., and I. Fujinaga. 2005. Automatic music classification and the importance of instrument identification. Proceedings of the Conference on Interdisciplinary Musicology. (This study used a previous version of jSymbolic called Bodhidharma.)