Digital Signal Processing Reference
In-Depth Information
6.2.5.
Bit rate constraint
For every 20 ms, the following parameters are transmitted:
- The coefficients for the filter
A
(
z
). In the LPC-10 coder,
P
=10.Taking
approximately 3 or 4 bits for a code coefficient, we find a bit rate of the order of
1.8 kb/s.
- The power
σ
Y
in the case of an unvoiced sound or in the same way for the
coefficient
α
in the case of a voiced sound which costs 50
×
6 = 300 bit/s (to cover
50 dB in 1 dB steps).
- The voiced/unvoiced distinction is 50 bit/s.
- Finally, the fundamental period 50
log
2
(
T
max
T
min
0
×
−
) = 350 bits/s.
0
Summing up all these contributions, an overall bit rate of 2.5 kb/s is obtained!
6.3. The CELP coder
6.3.1.
Introduction
Most telephone band speech coders, especially for bit rates in the range
4.8-16 kb/s, are
Code Excited Linear Predictive
(CELP) coders. The operating
principle of these coders was introduced by Atal in the beginning of the 1980s.
Atal's first basic idea [ATA 82] was to propose a new model for excitation of the
synthesis filter in the form:
K
y
(
n
)=
g
k
λ
(
n
−
n
k
)
k
=1
and to calculate the
g
k
and
n
k
parameters of this model while minimizing the criterion:
N−
1
x
(
n
)]
2
[
x
(
n
)
−
n
=0
There are two new features in this method when compared with determining the
input for the LCP-10 coder synthesis filter for voiced sounds. We are no longer dealing
with a comb because the impulse positions are no longer regular, and we no longer
render
S
X
(
f
)
x
(
n
). This is known as modeling
by synthesis
. At the transmitter, the output that will be realized at the receiver is
constructed explicitly.
≈
S
X
(
f
) but instead
x
(
n
)
≈
The second innovation by Atal, three years later [SCH 85], was to make use of
vector quantization which had just been developed. This idea involves suggesting an