This paper presents the application of a radically novel approach, called the Microcanonical Multiscale Formalism (MMF) to speech analysis. MMF is based on precise estimation of local scaling parameters that describe the inter-scale correlations at each point in the signal domain and provides e cient means for studying local non-linear dynamics of complex signals. In this paper we introduce an e cient way for estimation of these parameters and then, we show that they convey relevant information about local dynamics of the speech signal that can be used for the task of phonetic segmentation. We thus develop a two-stage segmentation algorithm: for the first step, we introduce a new dynamic programming technique to e ciently generate an initial list of phoneme-boundary candidates and in the second step, we use hypothesis testing to refine the initial list of candidates. We present extensive experiments on the full TIMIT database. The results show that our algorithm is significantly more accurate than state-of-the-art ones.