Taiscéalaí: Information Retrieval from an Archive of Spoken Radio News

In this paper we describe Taiscealai, a web-based system which provides content-based retrieval on an up-to-date archive of RTE radio news bulletins. Taiscealai automatically records and indexes news bulletins twice daily using a stream of phones recognised from the raw audio data. A user's typed query is matched against fixed length windows from the broadcasts. A user interface allows the news bulletins most likely to be relevant to be presented and the user to select sub-parts of the bulletins to be played. Many of the parameters we have chosen to use such as the size and amount of overlap of windows and the weighting of phones within those windows, have been determined within the framework of the TREC Spoken Document Retrieval track and are thus well-founded. We conclude the paper with a walkthrough of a worked example retrieval and an outline of our plans for extending Taiscealai into an integrated digital library for news.