Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 128

Table 5.12 Performance comparison of JQ-MSVQ and
MA-JQ-MSVQ

JQ-MSVQ MA-JQ-MSVQ

Number of bits per 20ms 36

Number of bits per set 18

Stages 6

Bit allocation 6,6,6,6,6,6

M 32

WMSE 2.227 e-04 1.622 e-04

Average SD (dB) 1.0926 0.9335

Outliers at 2 dB (%) 1.8135 0.5627

Outliers at 4 dB (%) 0.0052 0.0052

variations which violate the Nyquist sampling criterion. Therefore the use
of an anti-aliasing filter with cut-off frequency adequate for the chosen LSF
sampling rate may be used to alleviate possible spectral overlapping of
the LSFs. It is confirmed [15] that this method offers an advantage over
the classic LSF extraction methods; during quantization, bit-saving and
significant reduction in the percentage of outliers have been possible.

5.10.1 LSFExtraction
Al-Naimi investigated the speech stationarity assumption over the analysis
window with regard to LSF vector extraction by calculating LSF vectors at
every sample [15]. The centre of the analysis window is shifted by one sample
at a time, leading to an LSF vector extraction rate of 8 kHz. Evolution of each
LSF parameter over time, also referred to as an LSF track, is then produced
from the over-sampled LSF vectors. Decimation without any filtering of
the LSF tracks at a given LSF vector transmission rate (i.e. 20, 10 or 5ms)
should produce exactly the same LSF vectors as the classic methods. It is
therefore clear that LSF track frequency variations greater than half of the
LSF computation rate (frequency) will cause problems during the decimation
process i.e. by introducing aliasing distortion. Note that the LSF computation
rate need not be same as the frame transmission rate.
In order to measure the amount of aliasing introduced, the following test
was used:

1. Ten LPC parameters were calculated for every sample using Hamming
windowing over 200 samples and bandwidth expanded by 15Hz, then
converted to LSFs.
2. The evolution of eachLSF track f i over timewas taken and FFT transformed.
The logarithmic magnitude of the FFT spectrum is shown in Figure 5.18.