| Summary: | Labiodental fricatives are featured in many of the world's languages. Roughly 40% of languages have the voiceless labiodental fricative /f/ in their phonetic inventory (Maddieson, 2005). This research is focused on the relationship between speakers' oral anatomical structures and their articulation of labiodental fricatives. Data from the ArtiVarK dataset, a multi-ethnic database including intra-oral scan measurements, audio recordings, and midsagittal MRI video recordings, was used. It was hypothesised that the configuration of the upper incisors, overjet, and overbite were the most important causes for differences in articulation. The anterior oral anatomy was traced in 430 midsagittal frames of the MRI videos. This was used as training input for a convolutional neural network, which automatically traced the anterior oral anatomy of the rest of the videos. Principal component analysis was applied to the traces to parse out differences in articulation methods. VIF analysis and LASSO regression was conducted over 36 variables. Results from the VIF analysis supported the hypothesis while results from LASSO regression suggested that overbite, overjet, and incisor angles were not salient. Both analyses found that gender was important while palate height and oral dimensions near the canine area contributed significantly less to articulation than other oral structures.
|
|---|