GENDER GATES FOR TELEPHONE-BASED AUTOMATIC SPEAKER RECOGNITION

The present work demonstrates a need for enhancing text-independent, t elephone based, automatic speaker recognition systems with a gender ga te. A range of gender gates and speech parameter types are proposed fo r this problem. These gates and parameters are also investigated in th e context of speech degraded by coding and reverberation. It is found that the performance of the most accurate gender gates and speech para meters is similar for uncoded, coded, and reverberated speech. However , the most accurate gender gates and speech parameter types differ sli ghtly across the three scenarios. The most robust all-round gender gat es consist of two Mahalanobis distance classifiers with fused outputs or pitch fused to the output of one such classifier. The best all-roun d speech parameters were reflection and Mel-based cepstrum coefficient s. (C) 1997 Academic Press.