Digital Signal Processing Reference
In-Depth Information
Figure 7.15 Basic structure of the vocal tract transfer function estimation based on
codebook approaches.
frequency range from, for example, 200 Hz to 3600 Hz. The basic structure of a code-
book approach utilizing a bandlimited and a broadband codebook is depicted in
Figure 7.15. In the depicted structure a mixed codebook approach, consisting of a nar-
rowband cepstral codebook and a broadband predictor codebook, is used.
As in the neural network approach the predictor coefficients can be transformed
into another feature space, such as line spectral frequencies or cepstral coefficients
(as depicted in Fig. 7.15), which might be more suitable for applying a cost function.
However, also for predictor coefficients well-suited cost functions exist. The likeli-
hood ratio distance measure, that is defined as
!
dV
,
ð
p
jA
i
,nb
(
e
jV
)
j
2
1
2
p
d
lhr
(
n
,
i
)
¼
2
1
(7
:
49)
jA
nb
(
e
jV
,
n
)
j
p
is sometimes applied for this application. The quantities
A
nb
(
e
jV
,
n
) and
A
i
,nb
(
e
jV
)
denote the narrowband spectral envelopes of the current frame and of the
i
th codebook
Search WWH ::
Custom Search