Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Discriminative Phonemes for Speaker Identification

Eluned S. Parris, Michael J. Carey

Ensigma Ltd., Chepstow, Gwent, UK

This paper describes experiments in speaker identification using the discriminatory ability of subword models based on phonemes. Speaker identification is performed using a free match technique where test speech is matched to a combination of speaker independent subword Hidden Markov Models (HMMs) and speaker dependent subword HMMs. Errors occur when the true speaker's phonemes are not consistently matched with the speaker dependent models or when other speakers match to the true speaker's models. Using Bayesian statistics we show that the discriminative power of a phoneme is related to the log likelihood ratio of the frequencies of true speaker to impostor recognitions. Reduction in error rates of 33% have been achieved over a simpler technique for a ten speaker identification task. The phonemes z, i, e, m? v, N, { and s were found to be the most discriminative. More generally front vowels, voiced fricatives and nasals give the best performance.

Full Paper

Bibliographic reference.  Parris, Eluned S. / Carey, Michael J. (1994): "Discriminative phonemes for speaker identification", In ICSLP-1994, 1843-1846.