The relevance of wavelet representation of melodic shape

Gissel Velarde, Tillman Weyde

Research output: Contribution to conference without publisher/journalConference abstract for conferenceResearchpeer-review

Abstract

Background
Shapes in melodies have been computationally modelled by gestalts metrics [1], neural networks [2] and statistical descriptors [3], among others. Wavelet coefficients provide an alternative and so far hardly explored representation, obtained by the decomposition of a signal into wavelet components at different scales and times [4]. Wavelet coefficients plotted into a scalogram provide a visual representation that lends itself to visual detection of patterns and hierarchical structures.

Research question
Do wavelet coefficients represent information that matches human cognition of
melodic shape?

Aim
The aim of the study is to explore the relation between human melody cognition and wavelet coefficients with a qualitative and a quantitative approach.

Summary of content
We represent a melody as a function of pitch over time and apply wavelet analysis with the Haar wavelet, essentially a short up-down movement. Our empirical results show that wavelet coefficients enable significantly better performance than pitch-time representation and segmentation in a melody-recognition model.

We discuss the relation between wavelet coefficients and the musical features of a melody, such as phrase contour and boundaries, in case studies with melodic analyses and scalograms. Possible implications for models of melody perception and cognition will be discussed in the light of a new experiment on wavelet models for melodic similarity, which we are currently working on.

Significance
We found clear evidence that wavelets coefficients capture cognitively relevant
aspects o f melodies. This opens up interesting new opportunities for software tools and visualisation, as well as for automatic music classification and indexing in music information retrieval.

References
[1] Buteau, C., & Mazzola, G. (2000). From Contour Similarity to Motivic Topologies. In Musicae Scientiae,4/2, 125-49.
[2] Weyde, T. (2002). Integrating Segmentation and Similarity in Melodic Analysis. In: Proceedings of the International Conference on Music Perception and Cognition 2002, pp 240-43, Sydney.
[3] Conklin, D. (2006). Melodic analysis with segment classes. Machine Learning, 65(2-3):349 -60.
[4] Mallat, S. (2009). A wavelet tour of signal processing. Academic Press, Third Edition.
Original languageEnglish
Publication date12 Jul 2012
Publication statusPublished - 12 Jul 2012
Externally publishedYes
EventMusic & Shape Conference - Senate House, University of London, London, United Kingdom
Duration: 12 Jul 201214 Jul 2012

Conference

ConferenceMusic & Shape Conference
LocationSenate House, University of London
CountryUnited Kingdom
CityLondon
Period12/07/201214/07/2012

Fingerprint Dive into the research topics of 'The relevance of wavelet representation of melodic shape'. Together they form a unique fingerprint.

Cite this