AI Seminar for Nov 7
Thomas Bartold
the AI lab.
AI Seminar
Tuesday, November 7, 4pm
ATL Large Conference Room
Speaker: Professor William Birmingham
Title: Automated Analysis of Tonal Music
Abstract:
The harmonic analysis of music by computer is a long-standing theme in the
computer-music research community. Harmonic analysis, specifically the
identification and spelling of a chord's name, is difficult to automate,
particularly when trying to achieve the high-performance of trained
musicians. Furthermore, while musicians analyze tonal music of all types, it
is difficult to create algorithms that have this generality. In this talk,
we describe two algorithms that together perform harmonic analysis on any
piece of tonal music with very high performance.
We divide harmonic analysis into two tasks. The first is identifying those
places in the music where the harmony changes; we call this task
segmentation. We have developed an algorithm for segmentation that makes
minimal commitments to (musical) context, thus gaining flexibility over the
types of music it can process. We show how to accomplish the segmentation
task in O(n) time using several heuristics.
The second task is giving each segment the proper name, quality, root, and
inversion; we call this task labeling. For this task, we use a scheme that
exploits pitch relationships that must hold among structures, thereby
reducing the need for knowledge (e.g., rules) tailored to a particular type
of music. In the case of block chords, where notes start and end
simultaneously, the harmonic changes are obvious, and so harmonic analysis
is relatively straightforward. Music, however, is not this simple: chords
are often arpeggiated (spread out over time), incompletely stated, and
interspersed with myriad non-harmonic tones (often stating the melody), in
forms such as trills and appoggiaturas. All of these things confound the
harmonic-analysis task.