Correlation fractal characterization of speech consonants

This paper presents computation algorithms for a characterization scheme of speech consonants using a fractal model, by assuming turbulent sources for consonant excitation. First, the scheme estimates the excitation from a consonant waveform using a linear predictive coding (LPC) excitation. Using Takens embedding theorem, the scheme treats the LPC excitation as a time-series observation of one variable in the assumed excitation dynamical system, and constructs several synthetic excitation attractors of the dynamical system, for various selected values of a construction parameter called embedding dimension. The resulting attractors can then be characterized using correlation fractal dimensions. A preliminary observation on 22 consonants shows encouraging results because every consonant results in a unique trend of fractal dimensions for different embedding dimensions and measurement scales.