Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 122

Table 5.9 Prediction gain vs MA predictor order
Order Optimum prediction parameters Prediction gain (dB)

1 0.65 2.97

2 0.85,0.43 4.13

3 0.85,0.60,0.35 4.61

4 0.9,0.7,0.45,0.2 4.84

be seen from Table 5.9 that the increase in prediction gain when increasing
the order of the MA predictor from 1 to 2 is 1.16 dB. An increase in prediction
gain of 1.87 dB can be achieved by increasing the order from 1 to 4, whereas
the increase from order 0 (no prediction) to 1 is nearly 3 dB. Although higher
order predictors help to increase the prediction, the degradation in speech
quality due to channel errors is expected. If the order is 1, 40ms of speech are
corrupted. With proper error concealment techniques, it is usually possible to
limit the distortion caused by the loss of LPC for 40ms to an acceptable level.
However, for higher prediction orders, 60ms or more are lost and the speech
degradation caused by such a loss is usually difficult to recover, affecting
the overall speech quality significantly. Therefore, for most applications with
a 20ms parameters update rate which involve a noisy channel, it is better
to use a first-order MA prediction. In case of shorter update rates, or very
low bit error conditions, higher order prediction can be used to improve the
MA prediction performance. In the following discussion, only first-order MA
prediction is considered.

5.9.3 PredictionFactorEstimation
Figure 5.15 shows that the best prediction gain for a first order MA is achieved
with a value of 0.65. Therefore it would be reasonable to assume that a
prediction factor of 0.65 will give the best performance in a first-order MA
quantizer. Indeed, such a value is used in some speech coders such as EFR [10].
However, this value has been derived using the assumption that the original
residual r n is close enough to the quantized residual r n that it can be used

instead to obtain the curve shown in Figure 5.15. In a practical quantizer, there
is no guarantee that this assumption will be true. Therefore the only way to
determine the optimal prediction factor is by training quantizers with various
prediction factors and comparing their performances. Various first-order MA
quantizers have been trained for values of α ranging from 0.3 to 0.7 in 0.05
steps, for a 20ms update rate. An MSVQ quantizer comprising three stages
of 8 bits each has been selected to quantize the residual, as it provides good
performance. The performance of these quantizers is plotted in Figure 5.16,
together with the performance of the quantizer without prediction. It can be
seen that the best overall performance of the quantizer is achieved for a value

ˆ