Proceedings

Evaluating Wasserstein GAN Discriminators as Models of Phonological Well-formedness judgments

---

Abstract

A long-standing goal of phonological theory is to explain how learners derive well-formedness judgments from exposure to speech. While computational models of phonology typically operate over discrete representations and rely on supervision, human learners acquire phonological generalizations from continuous acoustic input. This paper evaluates whether a neural network trained only on raw speech can nevertheless produce well-formedness judgments. We test the discriminator of a Wasserstein WaveGAN as an unsupervised model of phonological well-formedness, focusing on its ability to distinguish harmonic from disharmonic vowel patterns. Harmony-trained models assign higher scores to harmonic items, while disharmony-trained models assign lower scores, indicating that the discriminator internalizes a phonological generalization without explicit supervision. We also test whether imposing architectural bottlenecks on the discriminator’s output layer encourages more categorical or interpretable behavior. Although the magnitude of harmony effects increases with model capacity, tighter bottlenecks do not systematically sharpen categorical distinctions or yield interpretable individual output features. These results suggest that GAN discriminators can function as unsupervised learners of phonological well-formedness from raw speech, while also highlighting some limitations in their ability to learn reliable grammar-like acceptability judgments. 

Keywords: Computational Phonology, Unsupervised Learning, Phonetics, Deep Learning, Vowels, Well-Formedness

How to Cite:

Ferenc Segedin, B., (2026) “Evaluating Wasserstein GAN Discriminators as Models of Phonological Well-formedness judgments”, Proceedings of the Annual Meetings on Phonology 2(1). doi: https://doi.org/10.7275/amphonology.3698