Speak Human

Abstract

Linguistic typology as well as articualtory phonology currently lack a model of stable pronounciation beyond individual languages. Whilst IPA wellformedness checks exist, these do not actually model human speech but rather systematic concerns. In this talk I present an empirical and interpretable Phoneme-based Neural Model of Universal Phonology (PhoNeMoUPhon) based on a typologically diverse corpus of IPA transcriptions. I discuss its advantages over speech models and its future use for Historical Linguistics.