Pitch Estimation and Voiced–Unvoiced Classification of Speech - Digital Speech: Coding for Low Bit Rate Communication Systems - page 167

y

y y

CL 1

−CL −CL −CL

CL x CL x CL x

− CL − 1

y = clc(x)

y = cpl(x) y = sgn(x)

Figure 6.17 Clipper functions

x

;− CL ≥ x ≥ CL

= =

y clp(x) (6.30)

0 ;− CL > x < CL

1 ; x ≥ CL

y = sgn(x) = − 1 ;− CL ≥ x (6.31)

0 ;− CL > x < CL

The centre clipped signal y(n) is generally defined as:

y(n) = f [ s(n) ] (6.32)

The clipper function f [ . ] can be any one of the functions in equations
(6.29) - (6.31). For the autocorrelation method, the clipped autocorrelation
function R c (τ ) is defined as:

− −

N 1 N 1

R c (τ ) = y(n)y(n + τ) = f [ s(n) ] f [ s(n + τ) ] (6.33)

n = 0 n = 0

More generally, these two clipper functions can have any combination, e.g.

N − 1

R c (τ ) = f 1 [ s(n) ] f 2 [ s(n + τ) ] (6.34)

n = 0

A set of typical combinations of f 1 (n) and f 2 (n) are listed in Table 6.1. It has
been shown that[3, 16]:

•

For high-pitch speakers, the differences in performance scores between the
various clipping combinations are small and probably insignificant.

• For low-pitch speakers, fairly significant differences in performance scores

exist. Combination 1 in Table 6.1 gives the worst performance for all
utterances in this class. Combinations 4, 5, and 6 (those involving one
unprocessed component) are also poor in their overall performances.