Data-Driven Audio Feature Space Clustering for Automatic Sound Recognition in Radio Broadcast News

Abstract

In this paper we describe an automatic sound recognition scheme for radio broadcast news based on principal component clustering with respect to the discrimination ability of the principal components. Specifically, streams of broadcast news transmissions, labeled based on the audio event, are decomposed using a large set of audio descriptors and project into the principal component space. A data-driven algorithm clusters the relevance of the components. The component subspaces are used by sound type classifier. This methodology showed that the k-nearest neighbor and the artificial intelligent network provide good results. Also, this methodology showed that discarding unnecessary dimension works in favor on the outcome, as it hardly deteriorates the effectiveness of the algorithms.