Analysis by Synthesis LPC Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 246

does not require extra computation or bits. Encoding and decoding processes
of the codebook index in this algorithm have three possibilities: D

≥ L ;

L/ 2 D < L ;and D < L/ 2, assuming D min < L/ 2. The total phase positions

considered in each possibility can be calculated as follows:

≤

1. In the case of D L , there will be a single excitation pulse located in

the first position of the secondary excitation vector, hence, a possible L
phase positions will be considered. If the submultiple is also greater than
L , then the process stops. However, if D/ 2 < L ,then L

≥

D/ 2morephase
positions will be considered where the excitation vector will have an extra
pulse located at position D/ 2. Therefore, the total phase positions will
be 2 L

−

−

D/ 2.
2. In the case of L/ 2 D < L the secondary excitation vector will have two

pulses, placed at the first and D th positions. Therefore, the total phase
positions to be considered is D .If D/ 2

≤

D min then, a further D/ 2phase
positions will be searched giving a total phase positions of 3 D/ 2.
3. Finally, when D < L/ 2, the secondary excitation will have pulses at every
D samples starting from the first position, resulting in a possible D phase
positions. If, however, D/ 2

≥

D min then a further D/ 2 phase positions are
considered giving a total of 3 D/ 2.

≥

The above possibilities are indicated to the receiver by the fixed subframe
size and the decoded pitch predictor lag D .
In informal listening comparisons of VSE, centre-clipped Gaussian and
PAME, PAME produced the best result by making the overall speech sharper.
This was the result of the periodic excitation part of the secondary excitation
matching voiced speech faster and hence more accurately. This is illustrated
in Figure 7.24 where, the pitch of a voiced onset is better reproduced by the
pitch adaptive excitation. In this figure we can also see that PAME tracks
voice changes much faster. It must be noted, however, that the performance
of PAME can be affected if the pitch predictor lag is chosen wrongly in
the first place. Therefore, it is important that during the LTP search, the
correct lag or its integer multiples are selected. The dependency of the PAME
performance on the pitch predictor lag can be removed if all the possible lags
(in a subframe) and phase positions are exhaustively searched. This, however,
requires more index values to be coded in the adaptive part of the codebook.
In this case, a set of primary excitation vectors are formed by placing a unit
amplitude pulse at the start of the excitation buffer x , and then after every P
samples. P is varied from D min (the smallest possible pitch) to L (the subframe
size) to get all primary vectors. Whilst D min is related to the minimum pitch,
it may also be varied to enhance fidelity. Therefore, for each P , the primary