Utterance length is known to be an important factor in text-independent speaker verification. Here, we examine the effect of length on speaker separation for text-dependent verification. When both test data and templates are continuous-word strings, separation increases with length with apparent saturation at about 2 seconds. Simulating a possible practical scenario using continuous-word test data and templates formed by concatenating discrete words did not achieve useful separation for the longest tokens, apparently because of word-boundary effects. When test data and templates are discrete-word with matching scores accumulated over the maximum length, however, separation was good.
Keywords: Speaker verification, text-dependent verification, utterance length