Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 116

1.5e04

1.24

1.22

1.20

1.4e04

1.18

1.16

1.3e04

1.14

1.12

1.10

1.2e04

SS TS2 TS4 TS8 TS16 TS32 FS SS TS2 TS4 TS8 TS16 TS32 FS

Search strategy Search strategy

3.50

3.00

2.50

2.00

SS TS2 TS4 TS8 TS16 TS32 FS

Search strategy

Figure 5.12 Performance comparison of various search techniques

in multiply-adds per input vector is given in Table 5.8. It is to be noted that,
in the test, codebooks have been trained using the SS algorithm. Therefore,
they are only optimal for an SS search. Better performance for the TS and FS
cases can be obtained by using the same search in the training as the one used
during the operation of the quantizer. This is illustrated in Figure 5.13, where
WMSE, average SD and outliers at 2 dB are plotted for the original codebook
and the retrained codebooks, for SS and TS with values of M ranging from 2
to 32. Due to the very high complexity of the FS, it was not possible to fully
retrain the codebook using FS, although the results are expected to be similar
to that of TS with M = 32.

5.8.3 PerceptualWeightingTechniques
Several weighting techniques were described in Section 5.6.4. A good weight-
ing technique should give a distortion measure which is well correlated with
the spectral distortion measure, which is our reference here. For testing, we