Clustering audio clips by context-free description and affective ratings

In the absence of context, the process of listening to acoustic scenes results in deriving an explicit semantic description and an implicit assessment of its acoustic properties in terms of its affective value. In this work, we mainly exploit the relationship between context-free associations of audio clips containing unconstrained acoustic sources with their affective values for clustering. Using over two hundred clips from the BBC sound effects library, we present a novel, quantitative method to compare the clusters of audio clips obtained using its context-free description with the clusters obtained from their affective measures; namely valence, arousal and dominance. Our results indicate that comparing clusters across representations is a suitable approach to determine an appropriate number of clusters to index audio clips in an un-supervised manner. In this paper we present our findings and examples of the resulting clusters of audio clips.