Digital Signal Processing Reference
In-Depth Information
and
ρ (ω
0
)
is given by,
A
l
max
l
[
A
l
D
(
ω
l
−
2
A
l
K(ω
0
)
1
ρ
(
ω
0
)
=
kω
0
)
]
−
(8.6)
l
=
1
where
K (ω
0
)
=
π
/
ω
0
,
sin
2
π
ω
l
−
kω
0
ω
l
−
kω
0
≤
ω
0
2
ω
0
D
(
ω
l
−
kω
0
)
=
for
(8.7)
2
π
ω
l
−
kω
0
ω
0
and
D (ω
l
0otherwise.
The voicing level (probability),
L
v
(δ)
(i.e. the ratio of the voiced bandwidth
to the speech bandwidth, 0
−
kω
0
)
=
≤
L
v
(δ)
≤
1), is defined as,
1
δ>
13
dB
1
L
v
(δ)
=
9
(δ
−
4
)
4
dB
≤
δ
≤
13
dB
(8.8)
0
δ<
4
dB
The advantage of estimating the voicing for independent bands is that it essen-
tially removes the spectral tilt, i.e. all the components are equally weighted.
When the voicing is based on a single metric, i.e.
δ
, the large amplitudes
contribute more to the overall decision. If they have been corrupted by back-
ground noise, it may result in a large voicing error [2]. Therefore, the voicing
estimates based on independent bands are more robust against background
noise.
8.3.2 HarmonicAmplitudeEstimation
The harmonic coding algorithms require the spectral amplitudes of the
harmonics, which can be estimated in a number of ways.
Peak-picking of the Magnitude Spectrum
Harmonic amplitudes may be estimated by simple peak-picking of the
magnitude spectrum and searching for the largest peak in each harmonic
band. The peak amplitude value,
S
w
(m
k
)
should be normalized by a factor
depending on the window function used, as follows:
A
k
=
|
S
w
(m
k
)
|
κ
ω
0
2
<
2
π
kω
0
<
ω
0
2
for
−
N
m
k
−
and
k
=
1
,
2
, ... ,K
(8.9)
Search WWH ::
Custom Search