The Influence of Word Detection Variability on IR Performance in Automatic Audio Indexing of Course Lectures

This paper presents a study of the influence of acoustic variability on topic spotting performance in an application involving automatic indexing of course lectures. The application involves users formulating keyword queries to an indexing system which includes phone lattice based acoustic representations of audio material, a mechanism for keyword searching of a phone lattice, and a measure for evaluating the relevance of short segments of the audio files with respect to the query. The paper describes the results of a study where keyword detection performance for query terms provided by domain experts is related to the information retrieval (IR) performance obtained using a novel information retrieval measure. IR performance obtained from text transcriptions of the audio recordings and from the audio recordings themselves are presented.