A robust, segmental method for text independent speaker identification

Abstract

A robust, segmental method for text independent speaker identification is presented. Probability models are created from training data for each of the speakers of interest. The test sessions are then segmented and the statistics from each segment along with the models are used to compute scores for each speaker. Robust methods are described for combining the scores over all segments into one score. This process is carried out for each of the segment statistics which are combined to form a final score. Zero errors are obtained on a subset of the Switchboard corpus consisting of 24 speakers using six 60 second training sessions for each speaker and 97 thirty second tests.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>