Pitch Estimation and Voiced–Unvoiced Classification of Speech - Digital Speech: Coding for Low Bit Rate Communication Systems - page 160

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0

50 100 150

50 100 150 50 100 150

(c)

(a) (b)

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0

50 100 150 50 100 150 50 100 150

(d) Lag (samples) (e) Lag (samples) (f) Lag (samples)

Figure 6.9 Comparison of SS and SS-SA ( β = 0 . 25 ) for the high-, mid-, and low-

pitched signals shown in Figures 6.8a-c

where β is a weighting factor, 0 ≤ β ≤ 1, controlling the effect of SA. The

SS - SA becomes SS when β 0. Examples examining

the characteristics of ϕ (τ ) and ST (τ ) , are shown in Figure 6.9; the measured
value of each subfigure is normalized by each maximum value. The input
speech signals used in Figure 6.9 are the same as the ones used in the
analysis of the STA. For the high-pitched signals (short pitch periods) in
Figure 6.8a, the lag corresponding to the pitch period double has a strong
peak in Figure 6.9a, which seems to be even stronger than the peak at the
correct pitch. The SS - SA alleviates this problem as shown in Figure 6.9d
where the peak at the correct pitch lag becomes prominent in comparison
with other peaks. For the mid- and low-pitched (long pitch period) signals in
Figures 6.8b and 6.8c, the maximum peaks of ϕ (τ ) and ST (τ ) are relatively
obvious as illustrated in Figures 6.9b, 6.9c, 6.9e, and 6.9f.

= 1andSAwhen β =

Comparison
An objective test was conducted to determine various tuning factors. The
performance of the PDAs was measured in terms of pitch error rates ( E p ).
The speech test material, sampled at 8 kHz and filtered through the modified
intermediate response system (MIRS) [21], was composed of 56 seconds each
of male and female speech, each uttered by eight speakers. The reference
pitch periods were manually marked for each 10 ms frame.