Estimating vocal tract length by minimizing non-uniformity of cross-sectional area

Previous approaches to estimation of vocal tract length (VTL) differ in what information about the speaker is assumed to be known and which formants are treated as better predictors of VTL. However, they are alike in modeling formant frequencies as deviations from the resonances of a uniform tube, and in allocating equal credibility to all vowel spectra as predictors of length. The latter may be problematic, as phonotactic asymmetries in vowel quality in a data set can draw formants’ mean frequencies away from their underlying resonances, skewing VTL estimates. Herein, an additional parameter is proposed privileging vowel spectra that approximate the resonance characteristics of a uniform tube. The metric for this proximity is standard variance in Phi (SigmaPhi) across a vowel spectrum. In this study, formant data from 32 participants were analyzed using the estimators compared in Lammert & Narayanan (2015). Each estimator was run using frequencies from all vowel spectra, as well as from spectra with low values of SigmaPhi. As the threshold for maximum SigmaPhi was lowered, VTL estimates became tighter and the estimators typically converged on a length. This approach requires no labeling or identification of vowel type, and is therefore easily replicable across languages and speakers.