Digital Signal Processing Reference
In-Depth Information
Figure 7.12 Structure of the linear mapping approach for estimating the
broadband spectral envelope (using different parameter
sets
for voiced and
unvoiced utterances).
The classification of the type of sound can be performed by analyses such as the zero-
crossing rate [7] or the gradient index [20, 32]. Figure 7.12 shows the structure of the
linear mapping approach using cepstral coefficients.
Since the first analysis stage results in predictor coefficients a feature vector trans-
formation according to (7.21) is applied, resulting in a vector
c
nb
(
n
) containing the
narrowband linear predictive cepstral coefficients. After narrowband mean value sub-
traction, application of the mapping matrices, and broadband mean value addition the
resulting broadband cepstral vectors are transformed back into predictor coefficients.
This can be achieved also in a recursive manner according to
i
X
i
1
1
a
bb,
i
(
n
)
¼ c
bb,
i
(
n
)
kc
bb,
k
(
n
)
a
bb,
ik
(
n
),
k¼
1
for
i
[
{1,
...
,
N
bb,pre
}
:
N
bb,pre
denotes the length of the broadband predictor coefficients vector. Since stability
of the resulting IIR filter cannot be guaranteed any more, a stability check needs to be
Search WWH ::
Custom Search