Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 109

Table 5.4 Typical examples of SVQ LSF quantizers (24
bits/frame)
Sub-vectors Elements per subvector Bit allocation

2 5,5 12,12

3 3,3,4 8,8,8

4 3,2,2,3 6,6,6,6

5 2,2,2,2,2 5,5,5,5,4

Table 5.5 Complexity and memory requirements for various SVQ
schemes
Sub-vectors Split Bits Complexity Memory storage

10 8 10 8

1 10 24 1 . 67 × 1 . 67 ×

2 5,5 12,12 40 960 40 960

3 3,3,4 8,8,8 2560 2560

4 3,2,2,3 6,6,6,6 640 640

5 2,2,2,2,2 5,5,5,5,4 288 288

• The correlations between subvectors are not exploited. Therefore only a

fraction of the intra-frame correlation is used. In particular, a pair of LSFs
close to a peak in the spectrum may be split into two different subvectors
and, although there is a correlation between them, they are quantized
independently. As a result the quantization efficiency decreases greatly as
the size of the subvectors reduces.

• Some combinations of subvectors do not respect the ordering of the LSF,

or lead to neighbouring LSFs being too close to each other. As there is
a minimum spacing limit that a pair of adjacent LSFs are allowed to
have, this means that certain SVQ vector combinations will never be used,
which is a waste of bandwidth. This can however be alleviated to some
extent. For example, once the first subvector has been quantized, a simple
transformation such as an offset shift can be applied to the vectors that
violate the minimum distance in the second codebook, so as to make them
usable. However this is difficult to include in the training process, and the
resulting quantizer may not be optimal.

• The number of bits allocated to each subvector is fixed. The effect of the

weighting function will therefore be limited to within one subvector. If a
subvector contains only LSFs of relatively small importance, they will still
use all the bits allocated to this subvector, whereas a classic VQ would
effectively shift some of that bandwidth towards the more important LSFs,
through the weighting function. This effectively reduces the use of the
weighting function to the LSF within a given subvector and lowers the
overall quantization efficiency of an SVQ quantizer.