Digital Signal Processing Reference
In-Depth Information
100.0
Signal
Power
80.0
Total
60.0
STP
40.0
LT P
20.0
EXC
0.0
0.0
100.0 200.0
Time/40-sample subframe
300.0
Figure 7.16
Breakdown of SNR values for the CL3 LTP structure
codebook contribution. This is due to the fact that the codebook provides
contribution to match the remaining signal after the LPC and LTP contri-
butions have been subtracted from the original target. So the higher the
matching in LTP, the lower the contribution of the codebook will be. In
order to fill in the remaining information, the 10-bit codebook can only
provide up to a certain threshold in the overall SNR. In order to provide
more contribution, the LTP and codebook can be jointly optimized [19], or
a better codebook excitation source can be used.
Limited informal subjective listening tests have also shown that three-tap
LTPs are generally better than one-tap LTPs, although the difference between
the three-tap integer delay and the one-tap fractional delay (up-sampling
four times) is not very noticeable.
7.3.3 Multi-PulseExcitation
Early versions of CELP coding used multi-pulse excitation with and without
LTP to match the original input signal. As was briefly discussed earlier,
a low-complexity MPLPC coder sequentially determines the locations and
amplitudes of the excitation pulses so as to minimize the error between the
original and the synthesized speech. The optimum pulse locations are found
by computing the error for all possible pulse locations with their optimum
amplitudes in a given analysis block and selecting the allowable number of
locations and their amplitudes that result in the minimum error. A block
Search WWH ::
Custom Search