5.2. Parameterisation
107
Figure 5.10
Plot of a sample LPC spectrum with the corresponding LSP positions overlaid. Odd
lines are drawn solid and even lines are drawn dashed.
frequency lies somewhere between the open and closed model positions, represented
by odd and even lines). Local minima in the spectrum tend, by contrast, to be avoided
by nearby lines. These and several other properties explain the popularity of LSPs for
the analysis, classification and transmission of speech.
5.2.4.1
Derivation of LSPs
Line spectral frequencies are derived from the linear predictive coding (LPC) filter
representing vocal tract resonances in analysed speech as we have seen in Section 5.2.1.1.
For Pth-order analysis, the LPC polynomial would be:
A
p
(z) = 1 + a
1
z
−1
+ a
2
z
−2
+ · · · + a
P
z
−P
.
(5.20)
We will define two
(P + 1)th-order polynomials related to A
p
(z), named P(z) and Q(z).
These are referred to as antisymmetric (or inverse symmetric) and symmetric in turn
based on observation of their coefficients. The polynomials represent the interconnected
tube model of the human vocal tract and correspond respectively to complete closure
at the source end of the interconnected tubes and a complete opening, depending on an
extra term which extends the Pth-order polynomial to
(P + 1)th-order. In the original
model, the source end is the glottis, and is neither fully open nor fully closed during the
period of analysis, and thus the actual resonance conditions encoded in A
p
(z) are a linear
combination of the two boundaries. In fact this is simply stated:
A
p
(z) =
P
(z) + Q(z)
2
.
(5.21)
The two polynomials are created from the LPC polynomial with an extra feedback term
being positive to model energy reflection at a completely closed glottis, and negative to
model energy reflection at a completely open glottis:
108
Do'stlaringiz bilan baham: |