Towards Methods for Efficient Access to Spoken Content in the AMI Corpus

摘要

Increasing amounts of informal spoken content are being collected. This material does not have clearly defined document forms either in terms of structure or topical content, e.g. recordings of meetings, lectures and personal data sources. Automated search of this content poses challenges beyond retrieval of defined documents, including definition of search items and location of relevant content within them. While most existing work on speech search focused on clearly defined document units, in this paper we describe our initial investigation into search of meeting content using the AMI meeting collection. Manual and automated transcripts of meetings are first automatically segmented into topical units. A known-item search task is then performed using presentation slides from the meetings as search queries to locate relevant sections of the meetings. Query slides were selected corresponding to well recognised and poorly recognised spoken content, and randomly selected slides. Experimental results show that relevant items can be located with reasonable accuracy using a standard information retrieval approach, and that there is a clear relationship between automatic transcription accuracy and retrieval effectiveness.