Speaker verification performance and the length of test sentence

It is known that the performance of a speaker verification system improves with the length of test sentences. However, little is known about the exact relation between the performance and the test length. That makes it difficult to compare the results from various studies in which different test lengths have been used to evaluate the systems. In this paper, we have proposed a method to calculate the verification error rates at any lengths of test sentences, as long as the error rates at two different lengths are given. The accuracy of this calculation method is demonstrated with a speaker verification experiment and with the results reported in literature. Good agreement is shown between the calculated values and that measured through experiments.