Retrieving radio news broadcasts in Danish: accuracy and categorization of unrecognized words

Abstract

Digital archives of radio news broadcasts can possibly be made searchable by combining speech recognition with information retrieval. We explore this possibility for the retrieval of news broadcasts in Danish. An average of 84% of the words in the broadcasts was recognized. Most of the unrecognized words were compounds, names, and other words that appear of value to retrieval. Thus, the set of words describing a broadcast has to be expanded to compensate for the recognition errors. We discuss doing this by exploiting the alternative matches from the speech recognizer and by extracting words from a related corpus.