Towards a More Complete Computational Model of Phonation

Phonation contrasts have proven difficult for linguists to accurately and adequately categorize cross-linguistically. While particular acoustic features may robustly distinguish one phonation type from another in a given language, no known acoustic features robustly distinguish all phonation types in all known languages with phonemic phonation-type contrasts. This study presents an exploration of possible resolutions to the challenge of classifying phonation types in a cross-linguistically consistent manner. A novel adaptation of the ExSTraCS learning classifier system demonstrates some similarities between Black Miao modal voice and Yi lax voice. This model also demonstrated, surprisingly, that CPP was the most effective predictor of Black Miao and Yi phonation types, but that it loses effectiveness in the presence of low values of H1*. A second experiment utilizes the same model with a balanced dataset generated by the ROSE resampling algorithm. This experiment demonstrates that, while class imbalance does pose a challenge for the construction of models that can accurately classify phonation types, the primary challenge remains cross-linguistic inconsistency in the acoustic measures used to classify phonation. This study concludes by using the results from the previous two experiments to propose a new phonation-type paradigm which is designed to be as cross-linguistically consistent as possible while maintaining within-language phonemic distinctions.