Evaluating Phonological Knowledge in Foundational Speech Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Speech foundation models are primarily evaluated with task-level metrics, which do not establish whether languagespecific phonological patterns are encoded or faithfully reproduced.This dissertation asks what evidence is required to distinguish recoverable acoustic or distributional information from phonological generalization in speech models.Assamese Advanced Tongue Root (ATR) vowel harmony provides the primary test case because its production pattern combines acoustic vowel contrasts with long-distance, directional dependencies.Four interconnected studies examine (i) the acousticphonetic realization of ATR in human speech, (ii) generative learning with fiwGAN, (iii) process-level probing of frozen selfsupervised speech models, and (iv) phonological faithfulness in multilingual TTS.Results so far show that ATR-and harmonycorrelated information can be highly recoverable even when model behavior does not exhibit the structural signatures expected of Assamese harmony.The dissertation, therefore, develops evaluation criteria that distinguish phonetic recoverability from evidence for phonological process encoding, with implications for speech technologies in typologically diverse and low-resource languages.