Digital Signal Processing Reference
In-Depth Information
(a)
20.0
(b)
18.0
16.0
(c)
14.0
12.0
10.0
8.0
6.0
0
10
20
30
40
Number of pulses/160 samples
Figure 7.19
The number of pulses against the segmental SNR: (a) amplitude re-
optimization after each pulse, (b) amplitude re-optimization after all pulses, and (c)
sequential MPLPC with no amplitude re-optimization
Subjectively, curves
(a)
and
(b)
are superior only when a high number of
pulses (eight or more every 20ms) are employed in the process of amplitude
re-optimization. This is expected since the re-optimization process improves
the performance of closely placed pulses. On average, five pulses per 4-5ms
are adequate to achieve good speech quality. One major problem during
the search for the pulses is pulse-doubling. Pulse-doubling usually occurs in
voiced regions with greater than about eight pulses every 10ms and involves
the re-selection of already-selected pulse positions. In order to avoid this
effect, the newly-found pulse amplitude is added to the existing amplitude
or the already-found pulse locations are excluded from further pulse position
selection. If joint pulse amplitude re-optimization is applied every time a new
pulse is positioned, pulse-doubling is eliminated automatically.
MPLPC with Pitch Prediction
A basic multi-pulse coder produces satisfactory speech quality at medium
bit rates. However as the bit rate is lowered, degradations in the speech
quality become noticeable. This is especially true for the higher-pitched
voiced regions which usually occur with female speakers. This is due to a
limited number of pulses being available, the majority of which are used to
model the fundamental pitch pulses and hence relatively few pulses remain
for the modelling of the remaining excitation signal. With the introduction of
a pitch predictor into the AbS loop as shown in Figure 7.20 such effects can
be reduced. In 1989, Singhal and Atal [12] proposed a closed-loop solution
Search WWH ::
Custom Search