Encode Decode
Encode Decode
Encode Decode
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
CODE.PPT 6.9<br />
CODE.PPT 6.10<br />
Quantising a Gaussian Signal<br />
Adaptive Quantization<br />
a 1 a 2<br />
x 1 x 2 … … x n–1<br />
Input values from a i–1 to a i are converted to x i .<br />
For minimum error we must have a i = ½(x i + x i+1 )<br />
x n<br />
For this range of x, the quantisation error is q(x) = x i – x<br />
The mean square quantisation error is<br />
+∞<br />
∫<br />
n<br />
a<br />
2 2<br />
∑ ∫ ( i )<br />
ai−1<br />
i = 1<br />
i<br />
E = p( x) q ( x) dx = p( x)<br />
x − x dx<br />
−∞<br />
Differentiating w.r.t. x i : ∂E<br />
ai<br />
= px( x x)<br />
dx<br />
∂x<br />
∫ −2<br />
( ) −<br />
a<br />
i<br />
i−1<br />
i<br />
ai<br />
ai<br />
= 2x p( x) dx−2<br />
xp( x)<br />
dx<br />
Find minimum error by setting the derivative to zero<br />
x<br />
i<br />
=<br />
∫<br />
∫<br />
ai<br />
ai−1<br />
ai<br />
ai−1<br />
xp( x)<br />
dx<br />
pxdx ( )<br />
To find optimal {x i }, take an initial guess and iteratively use<br />
the above equation to improve their estimates.<br />
For 15 bins, SNR = 19.7dB<br />
i<br />
∫<br />
∫<br />
ai−1 ai−1<br />
⇔ x i is at the centroid of its bin.<br />
– We want to use smaller steps when the signal level is<br />
small<br />
– We can adjust the step sizes automatically according<br />
to the signal level:<br />
• If almost all samples correspond to the central few<br />
quantisation levels, we must reduce the step size<br />
• If many samples correspond to high quantisation levels,<br />
we must increase the step sizes<br />
Page 6. Speech Coding E.4.14 – Speech Processing