Finite element generation of sung vowels tuning 3D mri-based vocal tracts

Vowel sounds can be generated from the finite element simulation of acoustic waves propagating through a detailed three-dimensional (3D) vocal tract (VT). The latter can be obtained from Magnetic Resonance Imaging (MRI). However, current databases in literature are limited to spoken vowel sounds. In this work, we propose to tune (i.e. modify) spoken VT geometries to attain some of the acoustic effects that can be found in singing voice, such as the singing formant. To do so, the original 3D VT is first discretized into a set of cross-sections to obtain a 1D representation of it, by means of area functions. Secondly, the area functions get modified to obtain the desired formant frequencies following a tuning methodology based on sensitivity functions. The resulting 1D VT is finally reverted back to 3D combining information from the modified area functions and the original shape and location of each cross-section. Several finite element simulations are carried out with the resulting 3D tuned vocal tracts to illustrate the performance of the proposed methodology. An LF-model is used to generate a train of glottal pulses that includes the fundamental frequency fluctuations that characterize singing voice.