Tonal stability modeling from audio chroma features

This work explores the possibility of modeling tonal stability from music signals, in an attempt to approximate in a different way an ubiquitous, yet controversial, topic in music research: that of the perception of tension and relaxation over time. We rely upon some of the most agreed hypothesis about the hierarchical properties of the tonal system, to define a scenario in which such features could be modeled and exploited from audio signals. From the methodological point of view, we reinterpret current paradigms from music theory and cognitive psychology, in order to define a coherent and manageable concept of tonal stability, around the idea of perceptual degree of closeness between tonal events and their tonal contexts. Our proposed algorithm extracts chroma information from music signals and builds a representation in which several hierarchies of pitch events, related with traditional categories of pitches, chords and keys, are simultaneously accessible. The different levels on this keyscape establish the required hierarchical relationships between tonal events and -withintheir tonal contexts. This information is discussed in terms of stability through time and pitch space and, after its mapping into a geometric model of perceptual interkey distances, several metrics of tonal stability are proposed. Through three case studies, we discuss the applicability of the algorithm for cadence finding, tonal tension peaks modeling and stability-based structural analysis.