Speech Signal Analysis and Modelling - Digital Speech: Coding for Low Bit Rate Communication Systems - page 73

.

.

.

.

.

.

.

10

.

.

.

8

.

6

.

.

4

2

2 4 6 8 0 2 4 6 820 22 24

LPC Filter Order

Figure 4.8 LPCpredictiongainversusLPCorder

that the filter order needs to be at least eight. Usually, a 10-pole filter is used
so that formant resonances and general spectral shape is modelled accurately.
(However, much higher order filters have been used in elaborate schemes
such as the proposed ITU 16 kb/s G.728 standard [9], where a 50-order LPC
filter is used!). The LPC filter prediction gain versus the order of the filter is
shown in Figure 4.8 (note that the plot in this figure was obtained over a small
number of speech samples and it is only indicative, the LPC gain may vary
from sample to sample), and in Figure 4.9 the spectral envelope of various
filter orders is illustrated.
As for the frame size, the stationarity constraint applies, thus it is necessary
for us to choose a size which will conform to this. This usually implies
a frame size of 16-32ms. Another related factor is the partitioning points
of the analysis frame. The position of the analysis window may affect
significantly the performance of the LPC analysis. In order to reduce the
effect of window positioning some common preprocessing stages include the
use of pre-emphasis of the signal prior to the LPC analysis, and the use of
overlapping windowed frames. Overlapping frames try to overcome some
of the block-edge effects of the frame-based LPC analysis. The amount of
overlap is typically around 10-20 % of the frame size. Interpolation of the