Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

The 1999 NIST Speaker Recognition Evaluation, Using Summed Two-Channel Telephone Data for Speaker Detection and Speaker Tracking

Mark A. Przybocki, Alvin F. Martin

National Institute of Standards and Technology, Gaithersburg, MD, USA

The 1999 NIST Speaker Recognition Evaluation encompassed three tasks: one-speaker detection, two-speaker detection, and speaker tracking. All tasks were performed in the context of conversational telephone speech. The one-speaker task used single channel mu-law data; the other tasks used summed two-channel data. Twelve sites from the United States, Europe, and India participated in the evaluation. Performance was measured by a decision cost function and compared among systems and test conditions via DET Curves. Performance factors examined include segment duration, degradation resulting from the presence of a second speaker, sex mix of the two-speaker segments, matched or mismatched between training and test handsets, and the variation in handset type.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Przybocki, Mark A. / Martin, Alvin F. (1999): "The 1999 NIST speaker recognition evaluation, using summed two-channel telephone data for speaker detection and speaker tracking", In EUROSPEECH'99, 2215-2218.