Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 120

over a speech database. In order to increase the prediction accuracy, higher
order predictors can also be used. The prediction then becomes a weighted
sum of the LSF vectors for a given number of past frames. This increases the
performance of the predictor, at the small expense of slightly higher memory
requirements for storing the past values. Unfortunately this scheme has a
major drawback: the decoder must have correct knowledge of the prediction
used at the encoder. If a channel error occurs and corrupts the bitstream
for one frame, then the decoded LSF will be corrupted. Since the decoded
erroneous LSFs will be used for prediction, the LSF for the next frame will
also be corrupted and the error will then propagate indefinitely.
A better approach is to generate the prediction from the decoded codebook
entries, rather than the decoded LSFs which will limit error propagation.
Such predictors are called moving average (MA) predictors. A first-order MA
predictor is given by,

f n =

r k − 1 α n ˆ (5.77)

n

The decoded vector is then given by,

f n = ˆ

r n + α n ˆ r k − 1 (5.78)

n

Therefore if an error occurs, the only frames affected will be the frame where
the error occurs and the N following frames, where N is the order of the
predictor. For a first-order MA predictor, only one extra frame will be affected
compared with a quantizer not using prediction. Intuitively, an MA predictor
will not be as efficient as a DQ predictor, but its error resilience capabilities
are very significant. This makes the MA predictor a better choice for the
majority of applications.
Assuming all α n are chosen equal to a constant α , the prediction gains of
the DQ and MA predictors are plotted against α in Figure 5.15, for an update
rate of 20ms. Experiments show that forcing all α n to be equal does not
significantly reduce the prediction gain over the ideal case.
Figure 5.15 shows that the DQ predictor can achieve a gain of up to 5 dB
with an α value of 0.8, whereas the MA predictor can achieve 3 dB for α
around 0.65. The MA predictor is not as efficient as the DQ predictor, but still
provides a useful prediction gain, which in turn can help improve the overall
performance of the quantizer. The prediction gains for both DQ and MA
predictors depend on the LSF update rate which directly affects correlation
between adjacent sets of parameters. A faster update rate will give a higher
prediction gain, as consecutive sets of LSFs will be more correlated, and it will
usually be achieved with a higher value of α . For example, an update rate of
10ms gives an optimal α of around 0.8 for the MA predictor. In the following
sections, only the MA predictor will be considered as the DQ predictor is not