Automatic Text-Independent Syllable Segmentation Using Singularity Exponents And Rényi Entropy

Abstract

This paper presents an algorithm for syllable segmentation of speech signals based on the calculation of the singularity exponents in each point of the signal combined with Renyi entropy calculation. Renyi entropy, generalization of Shannon entropy, quantifies the degree of signal organization. We then hypothesize that this degree of organization differs which we view a segment containing a phoneme or syllable unit (obtained with singularity exponents). The proposed algorithm has three steps. Firstly, we extracted silence in the speech signal to obtain segments containing only speech. Secondly, relevant information from segments is obtained by examining the local distribution of calculated singularity exponents. Finally, Renyi entropy helps to exploit the voicing degree contained in each candidate syllable segment allowing the enhancement of the syllable boundary detection. Once evaluated, our algorithm produced a good performance with efficient results on two languages, i.e., the Fongbe (an African tonal language spoken especially in Benin, Togo, and Nigeria) and an American English. The overall accuracy of syllable boundaries was obtained on Fongbe dataset and validated subsequently on TIMIT dataset with a margin of error <5ms.