Sixth International Conference on Spoken Language Processing
October 16-20, 2000
Text-Independent Speaker Identification Using Gaussian Mixture Bigram Models
Wei-Ho Tsai (1,2), Chiwei Che (1), Wen-Whei Chang (2)
(1) Philips Research East Asia-Taipei, Taiwan
In this paper, a novel speaker modeling technique based on Gaussian
mixture bigram model (GMBM) is introduced and evaluated
for text-independent speaker identification (speaker-ID). GMBM
is a stochastic framework that explores the context or time dependency
of continuous observations from an information source.
In view of the fact that speech features are correlated between
successive frames, we attempt to investigate if speaker-ID can be
aided by modeling the spectral correlation in speech through the
usage of GMBMs. The proposed method was evaluated on a
100-speaker speech database. Experimental results demonstrated
that the error rate of speaker-ID could be greatly reduced by using
GMBMs, compared to the conventional speaker-ID technique
based on Gaussian mixture models (GMMs).
(2) Department of Communication Engineering,
Chiao Tung University, Hsinchu, Taiwan
Tsai, Wei-Ho / Che, Chiwei / Chang, Wen-Whei (2000):
"Text-independent speaker identification using Gaussian mixture bigram models",
In ICSLP-2000, vol.2, 314-317.