DefinePK

Abstract

While evaluating the performance of the speech perceptual
hash algorithms, we need to test their robustness, safety and real-time properties,
as well as judging the perceived similarity of the detected speeches. But the
existing algorithms are so insensitive to slight speech tampering that the tampered
speeches are mistakenly considered to be semantically unchanged. Therefore,
we present an algorithm for measuring the perceived similarity. By displaying
desirable sensitivity to slight speech tampering, the proposed algorithm can
detect slight speech tampering and judge whether the meanings have changed.
The proposed algorithm first divides the speech signals to many segments and
then performs correlation coefficient test on each segment in order to compute
the similarity. The experiment results show that the proposed algorithm can
effectively detect quality changes of the speech signals and the similarity
of the slightly tampered speeches. Its performance in evaluating perceived similarity
is superior to the popular similarity evaluation algorithms.

PDF
References