Browsing by Author "Sepúlveda Riquelme, Joaquín"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Enhanced prediction of model-based vocal features using a probabilistic bayesian neural network with uncertainty estimation
    (2024) Sepúlveda Riquelme, Joaquín; Cuadra Banderas, Patricio de la ; Zañartu, Matías; Mery Quiroz, Domingo Arturo; Espinoza, Víctor; Cienfuegos Carrasco, Rodrigo Alberto
    This thesis presents a thorough investigation into the modeling and estimation of vocal function variables, providing advances in the field of non-invasive ambulatory voice monitoring. Two related studies are involved in the thesis. One of the studies is aimed to resolve the discrepancies between the Triangular Body-Cover Model (TBCM) of vocal folds and clinical data. The study explores theimpact of physical properties, particularly the attenuation factor of the vocal tract, in accurately replicating clinical data, offering insights into the optimization of synthetic voice models for more accurate clinical representation. The second study introduces a novel application of a Probabilistic Bayesian Neural Network (PBNN) for estimating vocal function variables such as subglottal pressure,vocal fold contact pressure, and muscle activation variables, which are challenging to measure in ambulatory settings. The PBNN is trained on both synthetic and clinical data, demonstrating strong performance in predicting these variables with accurate es timations and narrow confidence intervals in synthetic contexts. In contrast, clinical contexts, incorporating transfer learning, present wider, more realistic confidence in tervals due to the inherent variability in human phonation. Furthermore, an observable correlation between prediction errors and both aleatoric and epistemic uncertaintieshighlights the ability of the network to forecast inaccuracies. Increased uncertainty at points of non-linear behavior, especially at higher subglottal pressures, suggests the need for improved input features to capture these nonlinear effects, indicating avenues for future research to enhance measurement fidelity.