Digital Signal Processing Reference
In-Depth Information
Figure 7.19 Branches of the spectral distortion measure for a¼ 0.1 and b0.0005.
Where
j
(
e
jV
,
n
) is defined as
<
D
(
e
jV
,
n
)
e
aD
(
e
jV
,
n
)
bV
,
if
D
(
e
jV
,
n
)
0,
j
(
e
jV
,
n
)
¼
(7
:
64)
:
ln[
D
(
e
jV
,
n
)
þ
1]
e
bV
,
else
:
Figure 7.19 shows the behavior of this distortion measure for different normalized
frequencies
V
. It is clearly visible that positive errors (indicating that the enhanced
signal is louder than the original one) leading to a larger distortion than negative
ones. Furthermore, errors at lower frequencies are leading also to a larger distortion
than errors at high frequencies. For computing the spectral distortion measure the
integral in (7.63) is replaced by a sum over a significantly large number of FFT
bins (e.g.
N
FFT
¼
256). Subjective tests as described in the next section show that
this measure shows a better correlation with human quality evaluations than stan-
dard distance measures. However, a variety of other distance measures that take
not only the spectral envelope into account might be applied, too.
7.6.2 Subjective Distance Measures
In order to evaluate the subjective quality of the extended signals subjective tests such
as mean-opinion-score (MOS) tests should be executed. As indicated at the beginning
Search WWH ::
Custom Search