Pitch Estimation and Voiced–Unvoiced Classification of Speech - Digital Speech: Coding for Low Bit Rate Communication Systems - page 165

12

10

8

6

4

STA / Male
STA / Female
SS-SA / Male
SS-SA / Female

2

0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Formant weighting factor ( )

γ

Figure 6.14 Analysis of the effect of the formant weighting factor γ in terms of the
pitch error rate; α and β defined in the STA and SS-SA functions are 0.5 and 0.25,
respectively, and the rectangular window is applied to TA calculation in the STA

intermediate signal between the original speech and its LPC residual signals.

A(z)
A(z/γ ) S(z)

S f (z) = (6.28)

where S(z) , S f (z) ,and A(z) are the z -transform of the input speech signal
s(n) , the formant-suppressed signal, and the inverse filter, respectively. The
parameter γ is the formant weighting factor, 0 1,

the filtered signal is identical to the original speech signal. On the other hand,
γ

≤ γ ≤ 1. For the case of γ =

0 makes the filtered signal equal to the LPC residual of s(n) .Itcanbe
seen that S f (z) is the intermediate spectrum between the original and residual
spectra for 0 <γ < 1. The effect of the formant weighing factor γ in equation
(6.28) was observed over the STA and SS - SA-based PDAs and the results
are shown in Figure 6.14. It can be seen that the value around 0 . 7

=

∼ 0 . 9gives

improved performance.
The effect of the flattening filter is shown in Figure 6.15. The formant
influence has been greatly reduced but not completely eliminated, while the
harmonic structure is well-preserved. A better performance may be obtained
by making the spectral-flattening factor a function of the average pitch
(tracked pitch) as shown in Figure 6.16.
Nonlinear spectrum-flattening is usually achieved by centre-clipping the
speech signal. The first centre-clipping PDA was proposed by Sondhi [22] in
1968 and various centre-clippers for autocorrelation PDAs were investigated
by Rabiner [3] in 1976. The characteristics of three types of centre-clipping