| Title |
Design Issues in Text-Independent Speaker Recognition Evaluation |
| Authors |
Martin Alvin (National Institute of Standards and Technology, 100 Bureau Drive, Stop 8940, Gaithersburg, MD, 20899, USA, alvin.martin@nist.gov)
Przybocki Mark (National Institute of Standards and Technology, 100 Bureau Drive, Stop 8940, Gaithersburg, MD, 20899, USA, mark.przybocki@nist.gov) |
| Keywords |
Evaluation, NIST, Speaker Recognition, Text-independence |
| Session |
Session SO6 - Recognition |
| Full Paper |
286.ps, 286.pdf |
| Abstract |
We discuss various considerations that have been involved in designing the past five annual NIST speaker recognition evaluations. These text-independent evaluations using conversational telephone speech have attracted state-of-the- art automatic systems from research sites around the world. The availability of appropriate data for sufficiently large test sets has been one key design consideration. There have also been variations in the specific task efinitions, the amount and type of training data provided, and the durations of the test segments. The microphone types of the handsets used, as well as the match or mismatch of training and test handsets, have been found to be important considerations that greatly affect system performance. |
