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 L <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">v</inf> . Monotonic increases in the probability of correct identification were obtained as L <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">v</inf> increased, 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 L <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">v</inf> corresponding 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.