IN THE PARAMETER SPACE FOR SPEAKER RECOGNITION
Hideki Noda · 1989
In forensic speaker recognition we would like to represent the result by speaker-dependent accuracy because there exist fairly large differences in accuracy among speakers. We believe that a recognition accuracy for an individual speaker depends on how variable and distinctive his utterance is. As for variation it is almost impossible to obtain enough samples of an utterance from a suspect to know about its variation. Our trials to estimate it indirectly have not yet succeeded for the present. The distinctiveness of an utterance may be defined as the distance between its position and the center of the distribution in the parameter space, which can be derived by mapping samples of the same utterance from many speakers. The use of this information has been found to be very effective for speaker recognition with speaker-dependent accuracy.