Content-based Retrieval from Unstructured Audio Databases Using an Ecological Acoustics Taxonomy

In this paper we describe a method to search for environmental sounds in unstructured databases with user-submitted material. The goal of the project is to facilitate the design of soundscapes in virtual environments. We analyze the use of a Support Vector Machine (SVM) as a learning algorithm to classify sounds according to a general sound events taxonomy based on ecological acoustics. In our experiments, we obtain accuracies above 80% using crossvalidation. Finally, we present a web prototype that integrates the classifier to rank sounds according to their relation to the taxonomy concepts.