Zero-crossing measurements for analysis and recognition of speech sounds

Abstract

Existing mathematical relations between the power spectral density and the mean zero-crossing rate of an ergodic random process are used to derive relations between spectral measurements and mean zero-crossing rates of a speech signal and its derivatives. Of particular significance is the equation relating zero-crossing rates to the formant parameters of vowels and vowel-like sounds. Reasonably close agreement between measured zero-crossing rates and those calculated from spectral measurements were observed for virtually all phonemes in a variety of contextual environments. Measured zero-crossing rates and corresponding calculated values for speech and its first derivative are presented for vowels, unvoiced fricatives, and unvoiced stops, all in many different contextual environments. Unvoiced fricatives /s/, / <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">\int</tex> /, and /f/ are shown to be distinguishable from each other solely on the basis of the zero-crossing rate of the derivative signal. Some vowels are shown to be differentiable from other vowels, although measurements other than zero-crossing rates of filtered, unfiltered, and differentiated speech are shown to be necessary for complete vowel separation. For unvoiced stop consonants, the zero-crossing rate of either the signal or its derivative is shown to be useful for classification, provided some information concerning the contextual environment is available.