Text-independent speaker identification and verification using the TIMIT database

This paper presents a neural network inspired approach to speaker recognition using speaker models constructed from full data sets. A similarity measure between data sets is used for text-independent speaker identification and verification. In order to reduce the computational effort in calculating the similarity measure, a fuzzy Vector Quantisation procedure is applied. This method has previously been successfully applied to a database of 108 Australian English speakers [1]. The purpose of this paper is to apply this method to a larger benchmark database of 630 speakers (TIMIT Database). Using the full 630-speaker database, an accuracy of 98.2% (one test sentence) and 99.7% (two test sentences) was achieved for textindependent speaker identification. On a 462-speaker subset of the database a 98.5% successful acceptance and 96.9% successful rejection rate for text-independent speaker verification was achieved.