Proceedings of the Annual Meeting of the Cognitive Science Society

What kinds of mental representations drive word recognition beyond surface phonetics? We address this question by studying the representations of a speech foundation model at two stages: 1) after pretraining on raw speech audio, and 2) after fine-tuning for word recognition. We focus on English word-final -er, which has three surface-identical morphological sources—comparative (bigger), agentive (diver), and monomorphemic (bitter). We find the model's representational space exhibits a linear geometric relationship between base and derived forms (big_bigger, dive_diver). This structure is specific to particular morphological sources, indicating that abstract morphology can be learned from unlabeled speech alone. However, the model overgeneralizes this pattern to false-friend pairs (bit_bitter). Fine-tuning for word recognition significantly strengthens this structure for true morphological relations, while suppressing relations in monomorphemes and lexically irrelevant information (ex., speaker identity). These results suggest abstract morpho-phonological structure is learnable from raw speech input, and plays a central role in word recognition.