Sculpting vowel space, an exploration

Abstract

This study investigates a method for generating synthetic speech that mimics the distorted vowel articulation found in speakers treated for oral cancer. Speech technology often performs poorly for these speakers, in part because little is known about how the articulatory consequences of oral cancer treatment impact speech intelligibility and quality, and because disordered speech data are scarce. To address this gap, we introduce a data augmentation method that directly manipulates vowel formant tracks, specifically F₁ and F₂, to recreate the characteristic flattening and reduction of vowel space associated with restricted tongue mobility after surgery. Using LPC based analysis-resynthesis, F₂ values are capped, F₂ dynamics are limited using low pass filtering, and F₁ is scaled to model compensatory jaw and lip movement. Manipulated vowel-like segments are then spliced into the original recordings while preserving non-vowel material to reduce synthesis artifacts. The resulting speech reproduces key acoustic properties of oral cancer vowel spaces, including severely reduced front-back distinction. Although artifacts remain and audio quality is imperfect, the method shows potential for creating controlled distortions useful for perceptual experiments and for expanding training data in inclusive speech technology development.