Analysis by Synthesis LPC Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 202

Note that, in the latest CELP coders, the above weighting filter has been
slightly changed by modifying its zeros (coefficients in the numerator) as well
as its poles.

A(z/β)
A(z/γ )

W(z) = (7.6)

p

a i β i z − i 1 −

i = 1

= 0 ≤ γ ≤ β ≤ 1

p

a i γ i z − i 1 −

i = 1

If this structure is to be used then in Figure 7.4, 1 /A w (z) should contain both
the above weighting filter and the usual LPC synthesis filter.

7.2.3 ExcitationSignal
The excitation signal represents the input to the AbS-LPC model and is
therefore an important block of the model shown in Figure 7.2. It provides
any residual structures that are not represented by the spectral model of
the time-varying filters, including pitch or long-term dependent structures
that exhibit significant correlation which is not covered by the pitch filter
and random structures that cannot be modelled efficiently by deterministic
methods. A proper excitation model is vital to the pitch-filtering efficiency
as the pitch filter memory is built up of its scaled versions. Therefore, the
excitation is usually represented by a shape vector with its associated gain or
scale factor. The various shapes that have been reported include multi-pulse,
regular-pulse, codebook, etc. Some mixtures of the above excitation schemes
have also been used.

Codebook Excitation
In codebook excitation (CELP) [10], the excitation vector is chosen from
a stored collection of C possible unity variance stochastic sequences with
an associated scaling or gain factor. In the AbS procedure, the C possible
sequences are systematically passed through the combined synthesis filter
(pitch, LPC and perceptual filters); the vector that produces the lowest error
is the desired sequence and is then scaled by its gain. Since the set of
sequences are present at both the encoder and the decoder, only an index to
the codebook and the gain level are required to be transmitted. Therefore,
less than 1 bit/sample coding is possible.
As the codebook is of finite dimension, it must be populated with represen-
tative vectors of the excitation. InAtal's original proposal, unit-variance white