Phonological complexity, speech style, and individual differences influence ASR performance for Tarifit

Abstract

Abstract This study examines individual differences through the lens of automatic speech recognition (ASR) transfer (applying ASR trained on one language to a new language) from Arabic to Tarifit, an under-resourced Amazigh language with typologically rare phonological structures. Thirty-seven native Tarifit speakers produced target words in both clear and casual speaking styles, allowing us to assess how phonological complexity and speech clarity interact to influence ASR performance. Results show that clear speech significantly improves recognition accuracy, particularly for words with rising sonority onset clusters. In contrast, falling sonority clusters and initial geminate consonants, which are both typologically marked structures, yield higher error rates even when spoken clearly. Importantly, we observe substantial speaker-level variability in ASR outcomes, though demographic factors such as age and gender do not predict performance. These findings suggest that individual differences in speech production and phonological encoding play a critical role in shaping ASR recognition success. By leveraging ASR as a proxy for perceptual processing, this work contributes to our understanding of how phonological structure and speaker variability jointly influence speech perception, with implications for inclusive ASR design and phonological theory.