Text-independent speaker identification from a large linguistically unconstrained time-spaced data base

A very large data base consisting of over thirty-six hours of linguistically unconstrained extemporaneous speech, from seventeen speakers, recorded over a period of more than three months, was analyzed to determine the effectiveness of long-term average features for speaker identification. The results were strongly dependent on the voiced speech averaging interval, or Lv. Monotonic increases in the probability of correct identification were obtained as Lvincreased, even with substantial time periods between successive sessions. Speaker identification performance in open tests improved if features with small between-class to within-class variance ratios were eliminated. For Lvcorresponding to approximately thirty-nine seconds of speech, true text-independent results (no linguistic constraints embedded into the data base) of 98.05% for speaker identification were obtained.