TU Berlin

Neural Information ProcessingMachine Learning and Neural Networks for the Perceptually Relevant Analysis of Music

Neuronale Informationsverarbeitung

Page Content

to Navigation

Machine Learning and Neural Networks for the Perceptually Relevant Analysis of Music


Music contains structural information as well as semantic connotations which are easy to perceive by human listeners, but which are difficult to extract automatically from an acoustic event (and even from the score of a given piece of music). Here we explore new techniques from the machine learning and the mathematical music theory fields with the goal to create semantically meaningful representations from acoustic events and to automatically extract perceptually relevant patterns from music and sound.

Acknowledgement: Research was funded by the EU and by the Technische Universität Berlin.

Selected Publications:

A Paradigmatic Approach to Extract the Melodic Structure of a Musical Piece
Citation key Adiloglu2006a
Author Adiloglu, K. and Noll, T. and Obermayer, K.
Pages 221 – 236
Year 2006
DOI 10.1080/09298210601045633
Journal Journal New Music Research
Volume 35
Abstract We present an automated, mathematical approach to the paradigmatic analysis of the melodic content of a piece of music. We consider all melodic segments of consecutive notes, however, segments of different sizes are processed separately. We compare and group these segments using a similarity measure, which accounts for standard symmetry transformations such as translation and inversion. We then define a significance measure for melodies via the number of repeats of a given melody and its close variations in the piece, and extract the melodies which appear more often than a threshold value. These melodies are then clustered, and – for every cluster – the melody which is repeated most often (including repeats of its close variations) is selected as the cluster's representative. After identifying the paradigmatic elements of each piece, we analyse them using the representative melodies found by the new method. In the present paper, we use a terminology inspired by topology, and indicate related links to it. We test our approach on the Two Part Inventions of Johann Sebastian Bach. We find that the representative melodies identified by our approach agree with the results of the traditional music theory well. Additionally, because the analysis is restricted to segments of consecutive notes, the implementation is fast and results can usually be analysed without the need for elaborate post processing.
Bibtex Type of Publication Selected:music
Link to original publication Download Bibtex entry


Quick Access

Schnellnavigation zur Seite über Nummerneingabe