Hierarchical Clustering Experiments for Application to Audio Event Detection

In previous work, it has been shown the feasibility of using an isolated sound effect corpus to train Audio Event Detectors (AED) for real life data. Thus, one can avoid the time-consuming task of manually annotating large amounts of movies, documentaries, TV shows or any other kind of data of interest. However, obtaining a quality sound effect corpus is still a tough task particularly when a large number of acoustic events is considered. In this case, unsupervised techniques able to classify semantic concepts can be very useful to avoid as much as possible the need for listening to the audio samples. In this paper, preliminary experiments involving hierarchical clustering of sound patterns are described. Both intraand interconcept clusterings have been carried out, at pattern level and at concept level by using means and variances of the set of selected features. Parameters of single mixture Gaussian models have also been used to identify audio concept similarities. Clusters of concepts with strong similarities between them have been obtained, both from the perceptual point of view and the semantic point of view. Additionally, these results are planned to be used to design an AED system that would have a hierarchical architecture.