Sixth International Conference on Spoken Language Processing
October 16-20, 2000
Bootstrapping for Speaker Recognition
Walter D. Andrews (1), Joseph P. Campbell (1), Douglas A. Reynolds (2)
(1) U.S. Department of Defense, (2) M.I.T. Lincoln Laboratory,
Lexington, MA, USA
The technique known as bootstrapping or resampling has been
used effectively in the field of statistics to obtain good estimates
of statistics from only a small set of observations. In this paper
we explore the use of this powerful technique to aid in
improving the performance of a GMM-UBM text-independent
speaker recognition system. We apply the bootstrap to the
training process in the generation of speaker models for the
GMM-UBM system. We also aggregate the outputs of the
bootstrap’s multiple speaker models in our bagging system.
Speaker recognition results of our bootstrap and bagging
systems are presented on NIST corpora.
Andrews, Walter D. / Campbell, Joseph P. / Reynolds, Douglas A. (2000):
"Bootstrapping for speaker recognition",
In ICSLP-2000, vol.2, 483-486.