Digital Signal Processing Reference
In-Depth Information
Fig. 11.3 Block diagram of the PCGMM method employing the proposed variational model
composition
where
r
e;k
is a constant bias term from the
k
th Gaussian component of the
e
th
environment model, and
p
is the posterior probability for environment
G
e
.
The variational noise models obtained by the proposed variational model com-
position method in this study are used to generate the environmental models {
G
e
},
which are estimated through the model combination procedure using the clean
speech GMM and the obtained variational noise models. A uniform prior probabil-
ity is set on all obtained noise models in this study. Figure
11.3
demonstrates the
resulting block diagram of PCGMM-based feature compensation employing the
proposed new variational model composition method.
ð
k
j
G
e
;
y
t
Þ
11.5 Experimental Results
As test data for performance evaluation, connected single digits portions from CU-
Move corpus were selected. We established an experiment setup which is identical
to the Aurora2 evaluation framework [
16
]. The task is connected English-digits
consisting of 11 words. Each whole word is represented by a continuous density
HMM with 16 states and three mixtures per state. In addition to the digits, two
silence models (i.e., normal silence and short pause) are used.
The feature extraction algorithm suggested by the European Telecommunication
Standards Institute (ETSI) was employed for the experiments [
17
]. The zeroth
cepstral coefficient was used instead of log energy for the sake of convenience in
model combination implementation. After extracting the 13th order cepstrum, the
first and second order time derivatives are included during the decoding procedure
(a total of 39 dimensional feature vector). The HMM parameters were estimated
using 8,840 clean speech training samples included in Aurora2, and performance
was evaluated on the selected test set of CU-Move corpus. The test set consists
of 464 utterances (length of 50 min) spoken by ten different speakers (five males
Search WWH ::
Custom Search