Application of the microcanonical multiscale formalism to segmentation of speech signals

In this paper we use a novel framework, the Microcanonical Multiscale Formalism (MMF), to analyze speech signals. The MMF is based on the computation of geometrical and local parameters — the singularity exponents — which allow nonlinear analysis of their complex dynamics and, particularly, characterize their intermittent signature. We define an accumulative measure on these exponents which has the nice property of producing clear and distinctive changes at phoneme boundaries. We present preliminary experiments on the TIMIT database, which show that singular exponents convey indeed valuable information about the local dynamics of speech. They also show that the measure we define has a good potential to provide a new and powerful method for text-independent phonetic segmentation.