Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 121

5.0

DQ prediction
MA prediction

0.0

5.0

0 0.2 0.4 0.6 0.8 1.0

Prediction Factor

Figure 5.15 Prediction gain of first order MA vs DQ Predictors (20ms update rate)

suitable for a general purpose coder (with possible channel errors). However,
a DQ predictor may be applicable in cases where virtually no channel errors
are encountered, such as voice storage applications.

5.9.2 PredictionOrder
MAprediction has been presented above for the case of a first-order predictor.
It is also possible to have a higher order MA predictor, where the prediction
is a weighted sum of the quantized residuals received in N previous frames.
An N th -order predictor would exploit correlations between the current frame
n and the frames n N . As a result, its performance is

expected to be greater than that of a first-order MA predictor. However,
the drawback of the improved performance is greater sensitivity to channel
errors as an error on one set of parameters will corrupt N

− 1, n − 2, ... , n −

1framesof
speech. The optimal order of an MA predictor requires a trade-off between
prediction gain and error resilience. In order to estimate the optimal order to
be used in most applications, where channel errors are expected, the optimal
prediction factors have been derived for various orders. This was achieved by
computing the prediction gain for all possible combinations of the prediction
parameters, in steps of 0.05, over a database of 30 000 LSF, extracted at 20ms
intervals with all silences removed. The results are shown in Table 5.9. It can

+