Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 126

Table 5.11 Comparative performance of JQ-MSVQ and
MA-MSVQ

JQ-MSVQ MA-MSVQ MA-MSVQ

Number of bits 44 15 18

Number of bits per 60ms 44 45 54

Stages 8 3 3

Bit allocation 6,6,6,6,6,6,6,2 5,5,5 6,6,6

M 32 32 32

Complexity (per 60ms) 374 400 62 400 124 800

Memory 13 560 960 1920

WMSE 1.541 e-04 2.574 e-04 1.594 e-04

Average SD (dB) 1.2576 1.6383 1.3053

Outliers at 2 dB (%) 4.6563 17.2014 4.3119

Outliers at 4 dB (%) 0.0 0.1159 0.0185

The performance of JQ-MSVQ is illustrated in Table 5.11. The LSF quantizer
used in the 1.2 kb/s coder, which quantizes three sets of LSFs jointly using
44 bits in an 8-stage JQ-MSVQ quantizer, is compared against a classic
MA-MSVQ quantizer of similar bit rate and one of similar performance.
Complexity and memory requirements are also indicated. In this example,
LSFs are extracted every 20ms. The results clearly show the advantage of
JQ-MSVQ over MA-MSVQ in terms of performance. JQ-MSVQ has the same
performance at 44 bits as MA-MSVQ at 54 bits, and is far superior to the
MA-MSVQ at 45 bits. Complexity is higher for the JQ-MSVQ, but this may be
reduced by lowering the depth of the tree search M . Memory requirements are
also higher for the JQ-MSVQ, but again they can be reduced by adding more
structure to the codebook (more stages of smaller sizes) and accepting a slight
reduction in performance. Overall, JQ-MSVQ is very effective at providing
reasonable LSF quantization at very low bit-rates. At 1.2 kb/s, only 72 bits are
available every 60ms for quantizing all speech parameters. Assuming that the
gain, pitch and voicing are quantized using 28 bits, only 44 bits are left for the
spectral parameters. As shown in Table 5.11, an MA-MSVQ quantizer would
not work well under those circumstances, giving significantly degraded
speech quality with over 17% outliers at 2 dB. However the use of JQ-MSVQ
quantization makes a 1.2 kb/s coder a practical possibility, with only 4.6%
outliers at 2 dB.

5.9.6 UseofMAPredictioninJointQuantization
When using JQ-MSVQ, the redundancies between the jointly quantized sets
of LSFs are exploited. Using MA prediction within the meta-frame will not
therefore achieve any more gain. Indeed, a JQ codebook using MA from