Skip to main content
Browse by:
GROUP

Data Seminar

Event Image
Thursday, October 23, 2014
11:45 am - 12:45 pm
Chris Tralie (Duke Univ.)

"Geometric and Topological Analysis of Musical Audio Data"Researchers in the Music Information Retrieval (MIR) community have traditionally focused heavily on methods derived from Fourier and wavelet analysis for applications such as machine music genre classification, cover song identification, and artist identification. In this work we show the power of interpreting musical audio sequences as geometric curves in high dimensions in addition to the more traditional analysis techniques, and we explore the added value of this representation. In particular, we show how summarizing a delay embedding of a raw audio signal with "community accepted features" (CAFs) based on the Short-Time Fourier Transform stabilizes these curves and allows both global musical phenomena, such as chorus/verse loops and bridges, and local musical phenomena, such as vibrato, beats, and clicks, to be described geometrically. We then show how tools from topological data analysis (TDA) can be used to quantify these geometric features on the high dimensional curves, and we devise learning schemes based on these features to improve music classification. This is joint work with Paul Bendich, Marshall Ratliff, Derrick Nowak, and John Harer. Please come and enjoy lots of fun music and animations during this talk. Also please visit an app I developed out of this work at http://www.loopditty.net to visualize any song on soundcloud.com using our technique.

Contact: Monique Brown