THALES_LIA_syst1 :

This system is our Baseline System based upon the ALIZE tool Kit
(obtained from a French government funding project and developed by
LIA), which corresponds to the LIA 2005 Speaker Verification
System. We have replaced in this system the front end processing
(acoustic features extraction and voice activity detector). The
feature extraction processing is based upon the ETSI standard AURORA
(ETSI ES 202 050 v1.1.4) fully dedicated to speech processing front
end. This system includes Mel-filter bank feature extraction, noise
reduction and blind channel equalisation. We use a feature vector
including 20 MFCC, 20 delta MFFC, delta Log Energy, 20 delta-delta
MFCC and delta-delta Log Energy (62 acoustic coefficients). Such an
analysis is performed each 10 ms. Voice activity detector (VAD) is
based upon a multi-band energy detector (multi-scale estimation of
energy pdf) combined with a temporal correlation detector. This VAD
could also maintain good tracking performances of high voiced frames
in the incoming speech stream. After Voice activity detection used
speech frames are voiced and represent 25% of the total amount of raw
speech material. The statistical speaker verification system uses 2048
Gaussians with TNORM statistical normalization.

CPU execution time (evaluation): near 26 hours with 1 CPU (P4 3ghz, 1
Gb RAM)