Towards automatic speaker retrieval for large multimedia archives
Marijn Huijbregts, David A. van Leeuwen · 2010
In this paper we discuss the challenges of scaling a speaker retrieval system for small audiovisual collections towards a speaker retrieval system for large audio (visual) archives. We show that with our large scale speaker diarization approach it is possible to perform query-by-example speaker retrieval; to search for audiovisual documents in which a particular person is talking. On a selection of the ICSI meeting corpus we obtain a Mean Average Precision of 0.49 and precision-at-ten of 0.70. On a much larger archive of three months of Dutch broadcast television we obtain a precision-at-ten of 0.52.