Digital Signal Processing Reference
In-Depth Information
Table 10.23
Age and gender baseline results obtained by SMO learnt pairwise SVM with linear
Kernel
Sub-Ch.
Task
UA
[
%
]
WA
[
%
]
Train versus develop
-
{
1
,...,
7
}
44.24
44.40
Age
{
1
,...,
7
}→{
C
,
Y
,
A
,
S
}
47.11
46.17
{
C
,
Y
,
A
,
S
}
46.22
45.85
Gender
{
1
,...,
7
}→{
x
,
f
,
m
}
77.28
84.60
{
x
,
f
,
m
}
76.99
86.76
Train
+
develop versus test
-
{
1
,...,
7
}
44.94
45.60
Age
{
1
,...,
7
}→{
C
,
Y
,
A
,
S
}
48.83
46.71
{
C
,
Y
,
A
,
S
}
48.91
46.24
Gender
{
1
,...,
7
}→{
x
,
f
,
m
}
81.21
84.81
{
x
,
f
,
m
}
80.42
86.26
Table 10.24
Selected speaker independent results for height (H) recognition on the TIMIT corpus
test partition; contextual information by feature inclusion of age (A), gender (G), American English
dialect (D), education level (E), race (R) or all of these (All)
Context CC MLE [cm]
- 0.296 7.05
R 0.286 7.09
G 0.299 7.01
A 0.314 6.94
A,G 0.317 6.91
A,R 0.302 7.00
G,R 0.290 7.05
A,G,R 0.304 6.98
All 0.306 7.07
CC, MLE for regression (speaker height in centimetres). 1 582 acoustic features, classification by
SVR with linear Kernel, SMO, complexity 0.05
Table
10.24
next depicts results of the speaker height assessment task in strict
speaker independence by employing TIMIT's training and test partitions as stated
above and exclusively adding speaker contextual meta-information by selected (pairs
of) supplementary traits as additional feature(s) to the acoustic vector. Given the
case of regression and a continuous ordinal task formulation, CC and MLE are the
measures of performance. Gains can be observed by gradual addition of ground-truth
supplementary speaker trait information aside of the target task. Little improvement
is found for the height recognition task by gender inclusion (1.2 % relative correlation
improvement), age inclusion (6.2 %) and combined age and gender inclusion (7.3 %)
with the latter being the only significant one. Interestingly, age inclusion helps more
for the assessment of height than gender inclusion, even though all speakers can
be assumed to have reached their maximal height given their ages above maturity.
