This paper compares 4 different ways of using initial ('training') recordings as reference data ('templates') to be compared with a test token for speaker verification. We show that merging the original training data into a single template - as in 2 of the 4 methods - incurs a loss in speaker discrimination relative to the remaining 2 methods which retain all the data. However, these techniques do have the advantage of reduced computational complexity in terms of both storage requirements and processing time for verification.

The best-performing of our methods used a decision rule in which 60% of the individual matches to each retained template had to satisfy a threshold criterion. However, statistical significance of the superiority of this implementation relative to the next-best, which took a mean of the 5 individual scores as a basis for its accept/reject decision, has not yet been established. This awaits further work with a larger database.