Analysis of Dynamic Filtering Properties for Vocal-Tract during Bangla Vowel and Vowel-Consonant-Vowel Sequence Production

Abstract

This thesis deals with the acoustic analysis of dynamic filtering properties for vocal-tract and the perceptional space of Bangla vowel. Here we capture Bangla vowel, vowel-consonant-vowel speech signal from native Bangla Speakers. Acoustic study of dynamic-shape of vocal-tract for Bangla vowels and vowel-consonant-vowel (VCV) sequences are carried out based on dispersion and cross-correlation of linear predictive coding (LPC) filtering coefficients and short-time Fourier transform (STFT) of formant trajectory. The standard deviation of LPC filtering coefficients and transitional energy of formant trajectories indicate the dynamics of vocal-tract. The consonantal-constriction in VCV sequence accelerates the vocal-tract transitional nature and the transitional nature yields lower-valued cross-correlation with more stable vowels. Fourier transform technique is utilized to determine the cross-correlation of two unequal length LPC trajectories. In the domain of vowel perception, speaker invariant principal components are determined from the multidimensional perceptual space for comparatively stable Bangla vowels. Linguistic content base multidimensional vector space can be formed using the formant frequencies and its dispersion related statistical moments and the well reorganization-able vowels may maintain a separable distance in the multidimensional space. In this work, perceptual vectors and their associated energies are evaluated by determining the principal components of the multidimensional space. Speaker invariance is numerically evaluated using the correlation with principal components of native Bangla speakers. In this research, an algorithm is also proposed for the determination of energy based on the redundancies of the canonical variates for the inter-speaker vowel consistency.