Prosodic word and phrase boundary detection based on F0 contour analysis using empirical mode decomposition

This paper describes a technique for detection of prosodic word and phrase boundary for Bangla language readout speech based on the Empirical mode of Decomposition(EMD) of Fcontour. In this method F₀ contour of the sentence is extracted using the open source software “Praat” and then a continuous F₀ contour is generated using interpolation. Empirical Mode of Decomposition operates on continuous logarithmic F₀ contour to decompose into a set of IMF(Intrinsic Mode Function) components. The sum of DC component and the component before DC gives the information about global variation or phrase component. It is observed that the IMF having the most energy gives the idea about accent component or local variation. In total 150 Bangla readout sentences, 724 lexical words out of which 526 prosodic words containing 137 two words together (in this case each prosodic word contains two lexical words or syntactic words) and 31 three words (in this case to form a prosodic word it contains three lexical or syntactic words) together, are analyzed in this study. The correct detection of prosodic word boundary from the onset time is within .091ms with 6% insertion errors. The results of EMD analysis are then compared with the Bangla grammar and Fujisaki model, which are satisfactory. With the help of these word and phrase boundary, this can be a way to analysis and synthesis for F₀ contour with the help of Fujisaki model parameters in future.