Enhanced Temporal Feature Integration in Audio Semantics via Alpha-Stable Modeling

Modern feature-based methodologies in semantic audio applications attempt to capture the temporal dependency of successive feature observations, which form the so-called texture windows. This paper proposes an enhancement of this type of processing, known as temporal feature integration, by employing and testing alternative deployable strategies. Specifically, data are fitted through commonly used statistical principles, estimating the parameters of a given probability density function that maximize the log-likelihood of the samples inside each texture window. The main statistical model that is set under investigation is the alpha-stable distribution because it can successfully represent signals, which the commonly used Gaussian curves fail to capture. Within this framework, the enhanced feature integration method is also elaborated, introducing new measures for feature modeling. The main objective of this work is to introduce an efficient feature engineering protocol for temporal integration, specifying a compact and robust set of aggregated audio parameters that can address the needs of many audio information retrieval systems.