The fourth order cumulant of speech signals applied to pitch estimation

Abstract EN

In a number of speech applications, such as coding, synthesis or recognition, it is crucial to make a reliable discrimination between voiced / unvoiced segments and accurately determine the pitch period.

The problem of an accurate estimation and decision in noisy condition remains open; Higher-order statistics (H.O.S) have inherent properties that make them well suited when dealing with a mixture of Gaussian and non-Gaussian processes.

This paper explores the fourth order cumulant using autoregressive model (AR(p)) and presents a new algorithm for pitch detection of voiced sounds with and without colored Gaussian noise and shows the superiority of the novel method over the classical methods such as cepstral method.