A novel approach to phonetic segmentation through local singularity analysis of speech

This paper presents the application of a radically novel approach, called the Microcanonical Multiscale Formalism (MMF) to speech analysis. MMF provides efficient means for studying local non-linear dynamics of complex signals. The formalism is based on precise estimation of local scaling parameters that describe the inter-scale correlations at each point in the signal domain. In this paper we introduce an efficient way of estimating these local parameters and then, we show that they convey instructive information about local dynamics of the speech signal that can be used for determination of phoneme boundaries. We thus develop a two-stage phonetic segmentation algorithm that uses these parameters to automatically detect the boundaries between phonemes. For the first step, we introduce a new dynamic programming technique to efficiently generate a list of phoneme boundary candidates from a piece-wise linear functional. In the second step, we use hypothesis testing to make the final decision. 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.