Analysis by Synthesis LPC Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 245

by the pitch prediction, resulting in a better system. However at low bit-rates
(increased vector sizes), during voiced onsets and transitions where the pitch
cannot build up fast enough to track the changes, the speech quality dete-
riorates significantly. The advantage of algebraic codebooks also reduces at
low bit-rates (i.e. at around 4.8 kb/s) as the number of pulse combinations
need to be severely restricted in order to allocate fewer bits for the secondary
excitation which results in distorted speech. Other important issue at low
bit-rates is the amount of noise added to speech from the secondary excitation
during steady state voiced regions. A constrained gain approach [27] helps to
produce cleaner voiced speech by limiting the power of secondary excitation
during steady state voiced regions. This section describes an adaptive code-
book excitation where the excitation pulse-positioning is made adaptive with
the pitch lag computed for the same subframe. This can be seen as a subset
of the algebraic codebook approach where the pulse positions are severely
restricted but made adaptive with respect to the pitch so as to increase their
chances of positioning them to locations where they are needed most.
In pitch adaptive mixed excitation (PAME), the static codebook is split into
two parts. The first part is made adaptive with respect to the pitch lag as
follows. The excitation buffer is filled with a unit sample amplitude every D
samples starting from the first location. The rest of the vector elements are set
to zero. During the search of the codebook, this vector is synthesized and its
phase position is determined by shifting its synthetic response one sample at
atimefor D

1 times. Each phase position is then treated as a new excitation
vector. In order to guard against pitch-doubling errors in the LTP search, if the
lag D is greater than 2 D min the same process is applied again by placing the
excitation pulses every D/ 2 samples. The total number of excitation vectors
searched is then found by adding the total phase positions considered. This
is similar to regular pulse excitation with the decimation factor of D and D/ 2.
After selecting the best excitation vector from the pitch-adaptive section of
the codebook using C a phase positions, the search continues in the second
part of the codebook which is fixed and contains centre-clipped overlapping
excitation. Here, a further C f

−

= −

C a vectors are searched and the best
performing vector index from the overall search process is transmitted to
the receiver. At the receiver, after decoding the pitch lag, the corresponding
excitation vector is decoded.
By forcing the secondary excitation to have pitch structure, it is possible
to match voiced onsets more accurately. This is because the pitch predictor
memory builds more quickly to track the incoming periodicity more accu-
rately and the secondary excitation provides the required periodicity where
the pitch predictor fails. This, of course, depends on the accurate computation
of the periodicity by the pitch predictor in the first place. Many other adapta-
tion schemes may be used to accurately place the secondary excitation pulses
every pitch period. The pitch predictor lag adaptation is useful because it

C