Digital Signal Processing Reference
In-Depth Information
14
f
g
12
10
8
6
0.35
0.45
0.55
0.65
0.75
Prediction Parameter
Figure 5.29
WMSE performance curves for a range of MA prediction parameters
new method always has a higher prediction gain compared to the classic
extraction and, as expected, the difference between them becomes smaller for
higher LSF update rates.
In the MA predictor used above, the prediction is a function of the unquan-
tized LSF prediction residual from the preceding set. As shown before,
when used in a quantizer such as MSVQ, the prediction will be a function
of the quantized LSF prediction residual and, hence, it is expected to be
different.
Figures 5.29-5.31 show the effect of quantizing the prediction residual on
the moving average prediction coefficient. These results were obtained by
varying the prediction coefficient in steps of 0.05 and training a multi-stage
VQwith three stages of 7 bits each withM-best factor of eight. Table 5.15 gives
a summary of the comparative performance results for the classic and new
methods of LSF extraction where the low-pass filtered method significantly
outperforms the classical method.
Search WWH ::
Custom Search