Sixth International Conference on Spoken Language Processing
October 16-20, 2000
A MAP Approach, with Synchronous Decoding and Unit-Based Normalization for Text-Dependent Speaker Verification
Johnny Mariethoz (1), Johan Lindberg (2), Frédéric Bimbot (3)
(1) IDIAP, Martigny, Switzerland
This paper presents an overview of some of the
research tracks in text-dependent speaker verification
followed by the Picasso Project (Research Work Package).
We focus successively on the training algorithm
(based on a MAP approach), the state-sequence
decoding procedure (by Synchronous Alignment) and
the score normalization (unit-based z-normalization).
On short utterances of 1 to 5 syllables
(Polyvar command words), we obtain an error rate of 5.6 % with
2 training sessions, which can be brought down to
3.3% with 3 additional sessions in incremental trainmg mode.
(2) KTH, Stockholm, Sweden
(3) IRISA, Rennes, France
Mariethoz, Johnny / Lindberg, Johan / Bimbot, Frédéric (2000):
"A MAP approach, with synchronous decoding and unit-based normalization for text-dependent speaker verification",
In ICSLP-2000, vol.2, 286-289.