Estimating Vocal Tract Length from Vowel Formants: A Multi-Method Approach using Acoustic Pharyngometry

This thesis explores the estimation of vocal tract length (VTL) based on formant frequencies extracted from 42 speakers in different vowel conditions. The estimations are compared to length measurements from Acoustic Pharyngometry (APh). The goal is to investigate the extraction of VTL from APh and the predictive power and accuracy of different estimation methods, including traditional and more complex regression models. For more than 400 thousand timestamps the first four formants were extracted and used for estimation per vowel. Estimation methods were compared based on the following metrics: correlation, error metrics and R2. Traditional models show a stronger correlation compared to regression models, up to 0.74, but they explain very little of the variance (0.06). Regression models explain more of the variance (up to 0.52), suggesting a better overall performance. A systematic bias (overestimation for short and underestimation for large VTL) was observed. Therefore, non-linear models were examined. They are more consistent overall with more models having R2 values around 0.4. But the systematic bias could not be explained. Finally, a two-step approach was proposed where the output of the first model serves as input for the second model for taking into account the bias. These findings highlight the potential of VTL estimation for systems as Automatic Speech Recognition or vowel normalization but show the limits and need for further validation as well.