Audio Features, Precomputed for Podcast Retrieval and Information Access Experiments

This paper describes how an existing collection of podcast material has been enriched with precomputed audio features. The feature set which is described in the paper is made available to facilitate more convenient information access experimentation to collections that include both audio and text data. A simple example analysis is given to demonstrate how the audio features can be used to score podcast segments for being entertaining, discussion oriented, or subjective, to fit the current TREC Podcast Track task.