of Microprocessors

SOURCE-SIGNAL ANALYSIS 585
again, the comparator goes negative. If the leakage current is low, the circuit
is capable of responding exclusively to the highest peak in the presence of
orher peaks of nearly equal height. However, too little leakage inhibits
response to decaying signals entirely. In a digital implementation, the precise
point of comparator switching can be determined by interpolation at the
switching point only, thus avoiding a significant increase in computation
elsewhere.
In an analog implementation, single-shot number 1 simply indicates
the beginning of a new period. The output converter is a period-to-voltage
converter followed by a reciprocal element. In a digital implementation, one
would simply count samples between peak detections for the period and do a
digital division or table lookup.
The adaptive control processor looks at the current frequency and sets
three detection parameters accordingly. The preprocessor filter cutoff is set at
about 1. 2 times the current frequency, while the blanking single shot is set
for 80% of the current period. The peak detector discharge rate is normally
set fairly high to allow tracking of decaying signals. During startup from
silence, however, it is set for little or no discharge. Startup also requires that
the low-pass filter revert to a high cutoff frequency and that the blanking
time be set to a low value. As soon as two peaks are detected, whi(Oh then
gives a preliminary period estimate, the parameters are set accordingly. If the
first estimate is erroneous due to multiple peaks, the filter cutoff and blanking
will continue to favor high frequencies until the peak detector has
charged to the absolute highest peak in the signal. At this point, a longer
period would be found and the feedback would adjust to favor it. The
processor also detects long dead periods and resets to the startup mode in
response.
One way to improve the performance of the detector is to construct a
duplicate that processes negative peaks. As before, if the two disagree, the
final pitch output should the lower of the two estimates. With this improvement,
about the only waveforms that it would consistently fail on
would be Fig. 16-17E and a backward version of Fig. 16-17F, the latter of
which is unlikely to occur unless the speech had been subjected to severe
dispersion.
The idea of using multiple detectors and taking a majority vote has
been carried to the limit in a scheme proposed by Gold and Rabiner. The
general idea is to make several (six) estimates of the period using individual
peak and valley (negative peak) amplitudes as well as differences between
peak and valley and peak and previous peak. These six estimates along with
the preceding two estimates for each are combined in a decision tree to obtain
the final estimate. If there is a significant lack in consistency among the 18
values, no decision is made and the sound is declared to be unpitched. A full
explanation of the algorithm is quite involved and would require too much
space to reproduce here. Although seemingly complex, the programming is
straightforward and efficient.