Pitch Estimation and Voiced–Unvoiced Classification of Speech - Digital Speech: Coding for Low Bit Rate Communication Systems - page 174

the above procedure and it is selected as the pitch estimate if it satisfies
the condition. This process continues until all submultiples have been tested
against this condition. If none of the submultiples of the initial pitch satisfy
the condition, then the initial τ 0 becomes the final pitch estimate.
In some cases, where the decision threshold is wrongly exceeded, a multiple
or submultiple of the correct pitch may be selected. This may cause significant
performance degradations. Therefore, when designing a vocoder that requires
accurate pitch estimation, other measures which can reduce the effects of these
occasional pitch errors should be considered.

6.3 Voiced-Unvoiced Classification

The voicing is another very important parameter which must be estimated
correctly for good quality speech reproduction. In the old vocoders, a single
(binary) voicing decision was made by classifying the frame (or half the
frame) as either voiced or unvoiced. However, it is well known that the
transitions are very important for good quality speech synthesis and most
of the time the transitions are a mixture of voiced and unvoiced signals.
Therefore a mixed decision of voicing has been developed and used in many
of the latest vocoders. In the following, we review and discuss both binary
(hard) and mix (soft) decision voicing.

6.3.1 Hard-DecisionVoicing
Voiced and unvoiced sounds have very well-known characteristics which can
be used to classify them reasonably correctly. Some of the most distinctive of
these characteristics are discussed below.

Periodic Similarity

The most prominent characteristic that separates voiced speech from unvoiced
speech is its regularity and fairly well-defined pitch. During voiced speech,
samples in one pitch period look very similar to the samples in the adjacent
pitch period. Hence, measuring the similarity between samples in consecutive
pitch cycles can give a reasonably good idea if the speech is voiced or
unvoiced. The measurement of similarity, Ps , can be computed by

N

T)

2

s(i)s(i −

i = 1

Ps = (6.42)

N N

s 2 (i) s 2 (i − T)

i = 1 i = 1