A speech fractal dimension study

Fractal dimension (D) quantifies the roughness of a temporal signal and est imates its degree of freedom, allowing a good approach of its fluctuations and roughness. Using a 16 kHz time sampling and the box-counting method, iv e studied Ds of some of the main French phonemes, i. e. [a], [e], [i], [o], [y], and consonants in a consonant-vowel context pronounced 4 times by 10 males and 10 females. For D measurement of long phonemes we used the dyadic Box Counting method and its 10 points D measurement (10pD). For plosion pa rt of plosives consonants, ive designed a Send Continuous Box Counting meth od devoted to D measurement of short single dimension temporal signal. In t he aim to approach infinitely small time scales, and to appreciate at least the tendency of these 10 points set, we calculated also the slope of the 3 last points (3pD). Our study consistently demonstrates that vowels are not fractal ; plosive consonants are not fractal ; long fricative consonants a re fractal ; males Ds are significantly higher females Ds, as far as only v owels and long voiced consonants are concerned; there is a significant diff erence (p < 0.01) between 3pD values of vowels (couple [a] [y] excepted), a nd fricative consonants (couple [integral]] [f] excepted). In case of nasal consonants, this categorisation is efficient using both 3pD and 10pD measu rements (p < 0.05). These results will be commented and discussed, in the a im of clinical use, i.e. dysphonia follow up and auditory prosthesis speech signal processing.