Speech perception and learning in the geometric typology of phonological inventories – GEOMPHON

The sound inventories of languages (consonant and vowel sounds used in a language to make words) are not arbitrary. Across the world's languages, sound inventories tend to be coherent: having one sound increases or decreases the chances of having certain other sounds, depending on shared properties of those sounds. As a result, when sound inventories are visualized as points in space, they tend to form geometric shapes that could not arise by chance; there is currently no explanation of this. We conduct behavioural tests to evaluate two kinds of explanations (perceptual explanations, and explanations grounded in abstract cognitive constraints on class formation). We evaluate computational models of early phonetic acquisition based on deep neural networks and large speech corpora, which assume that the perceptual explanation is correct, and see if they show improved performance. We also build models that allow us to reconstruct the sounds of inventories in earlier stages of languages based on present-day acoustic data, in order to help better study how these geometric shapes arise over time. By examining the geometry of sound inventories from these various angles over a period of two years, we hope to quickly narrow in on plausible hypotheses about the cause of this mysterious typological tendency, laying the way for future research.