Novel Method of Nonlinear Symbolic Dynamics for Semantic Analysis of Auditory Scenes

Discovering semantic information from complexsignals is a task concerned with connecting humans'perceptions and/or intentions with the signal's content. In thecase of audio textures from environmental sounds produced bya cheering crowd, laughter, crackling fire, car crash orexplosion, complex meanings of the events are inferred andappraised in a listener's mind, which triggers an affectiveresponse that is relevant for well-being and survival. In thispaper we contribute to the ongoing research on affectivesemantics from sound by proposing a novel learningframework of affective auditory scene analysis using arecently developed method of non-linear dynamic signalanalysis. Using an adaptive symbolization process that findsthe best audio structure representation in terms of thesymbolized sequence recurrence properties, we show thatmeasures of periodicity and complexity derived from ourmodel are relevant for the characterization of affect in auditoryscenes, and that they will perform better than state-of-the-artmethods relying on low-level acoustic features.