Encode Decode

CODE.PPT 6.17
CODE.PPT 6.18
Filter Adaptation
Prediction Filter
We adjust the b i to reduce the signal d 2 (n) by moving them
∂
∂
a little in the direction − ( ) =−
∂b d 2
n d n dn ( )
( ) 2 ( )
i
∂bi
Filters:
−1
−2
Az () = az 1 + az 2
We express D in terms of previous values of D:
−1
−2
−3
−4
−5
−6
Bz () = bz 1 + bz 2 + bz 3 + bz 4 + bz 5 + bz 6
Note that ∂ D S Q X S Q A +
= + − = + −
B A ∂B
i
= = z
− 1−
A D
∂ai
∂b
∂D
i
−i
1
z
Note that A(z) and B(z) filters involve only past values
b A D z −i
D ∂dn
( )
=− ≈− ⇒ ≈−dn
( −i)
∂ i 1−
∂bi
Signals:
1 B
Y = D + BD + AY ⇒ Y = + 1−
A D
X BD AY Y D A +
= + = − =
B 1−
A D
D= S+ Q− X = S+ Q− Y+ D ⇒ Y = S+
Q
In the absence of transmission errors, the y(n) output at the
receiver and transmitter will be identical and equal to the
input speech plus the quantisation noise
Hence we want to adjust b i in the direction d(n)d(n-i):
The function sgn() takes the sign of its argument. The
0.996 leakage factor assists in recovery from transmission
errors and prevents b i from exceeding ±2.
( )
b ← 0. 996b + 0. 008× sgn d( n) d( n−i)
i
i
Adjustment of a i is similar but more ad-hoc:
a ← 0.996a
+ 0.012×
sgn
1
2
1
a ← 0.996a
2
+ 0.008×
sgn
( v(
n)
v(
n −1)
)
( v(
n)
v(
n − 2) ) − 0.03a
× sgn( v(
n)
v(
n −1)
)
1
Page 6. Speech Coding E.4.14 – Speech Processing