ESCA Workshop on Automatic Speaker Recognition, Identification, and Verification

Martigny, Switzerland
April 7-9, 1994

An Evaluation of Supervised and Unsupervised Classifiers for Speaker Recognition

Kevin R. Farrell, Richard J. Mammone

CAIP Center, Rutgers University Piscataway, NJ, USA

An evaluation of supervised and unsupervised classifiers is performed for text-independent speaker recognition. Speaker models that use unsupervised classifiers are trained with only the data for that speaker whereas those using supervised classifiers are trained with the data for all speakers. The unsupervised and supervised classifiers considered here are the vector quantization (VQ) classifier and modified neural tree network (MNTN), respectively. The VQ classifier and MNTN are evaluated for several text-independent speaker recognition tasks within a 100 speaker corpus. For closed-set speaker identification the VQ classifier and MNTN demonstrate comparable performance. For speaker verification and open-set speaker identification, the MNTN consistently outperforms the VQ classifier both with and without cohort normalised scores.

Full Paper

Bibliographic reference.  Farrell, Kevin R. / Mammone, Richard J. (1994): "An evaluation of supervised and unsupervised classifiers for speaker recognition", In ASRIV-1994, 67-70.