Estimation of Vocal Tract Resonances Using Spectral Prominent Regions and Artificial Neural Networks

Abstract

Formants (peaks in the spectrum) are typically found in vowel regions. However, resonances of vocal tract are present in both voiced and unvoiced regions. It is important to focus on extracting vocal tract resonances from speech signal to provide compact and effective representation. This work applies multilayered neural networks for the estimation of vocal tract resonances from linear prediction cepstral coefficients. To train such a neural network, the voiced regions of speech signal are first probed for reliable spectral prominent regions. A sequence of spectral peaks detected by imposing continuity constraints in voiced regions is referred to as reliable spectral prominent regions. A multilayer feedforward neural network is then trained to model a mapping function from linear prediction cepstral coefficients to vocal tract resonances estimated in these reliable spectral prominent regions. This modeled function is used to estimate vocal tract resonances in all speech regions including unvoiced sections. Improvements in vocal tract resonances estimated by this modeled function are 37, 35 and 41%, respectively, for the first three resonances when compared to WaveSurfer, and 49, 32 and 29% for the first three resonances when compared to PRAAT.