Analysis by Synthesis LPC Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 233

pulse rates. At high pulse rates, the subjective difference between the MPLPC
with and without LTP saturates, since at higher rates the pulses from the
multi-pulse excitation can model the fundamental pitch pulses accurately.

Pulse Position Coding

The coding of pulse positions is usually performed by enumerative source
coding techniques [6]. The number of possibilities for placing M pulses in a
subframe of L samplesisgivenby,

L
M

L !
M ! (L

= = (7.61)

−

M) !

Hence the minimum number of bits required for coding these positions is

B min = ( log 2 ) (7.62)

where is the nearest integer greater than or equal to .Suchmethods

of pulse-position coding can be considered as a vector quantization of the
pulse positions. These techniques are not very favourable in bad channel
conditions and alternative coding methods are pursued. Another method is
the independent coding of the pulse positions. Although this method leads
to a higher bit rate, it is more robust to channel errors. The number of bits
required for independent pulse position coding is given by,

.

B min = M × ( log 2 L) (7.63)

Figure 7.22 shows the comparison between combinational and independent
coding of pulse positions for L 32. Clearly, combinational coding is by far

the more efficient at very high pulse rates, but at very low number of pulses
the difference is small (i.e. 3 bits). At such coding rates, the disadvantage of
an extra few bits per analysis frame for independent coding is reflected in the
complexity and the coder robustness.

=

Pulse Amplitude Coding

Efficient normalization is necessary for coding of the pulse amplitudes,
because the pulse amplitudes have a large dynamic range and direct quan-
tization requires a large number of bits. Normalization can be carried out
by the rms of the amplitudes. In such methods, the rms value must also
be included in the transmission. This inevitably leads to higher bit rates.
In most MPLPC designs, quantization of the first pulse is accomplished by
incorporating a large number of nonuniform quantizer levels (usually five
bits or more) and the rest of the pulses in the subframe are normalized with