Beyond the binary: Investigating gradient palatalization in Russian with deep learning models

Recent research demonstrates that posterior probabilities of phonological features derived from deep learning models, such as Phonet (Vásquez-Correa et al., 2019), reliably quantify phonetic variation, with applications in linguistic analyses (e.g., Spanish lenition) and clinical diagnostics (e.g., distinguishing Parkinson’s Disease from Atypical Parkinsonism; Wayland et al., 2024, 2023). Building on this framework, our study examines gradience in Russian secondary palatalization, traditionally characterized as a binary distinction between palatalized and non-palatalized consonants. However, emerging research challenges this binary view, revealing context-dependent gradient variation (Parker, 2015), with articulatory evidence supporting incomplete neutralization of the plain-palatal contrast in Russian (Oh et al., 2024). Using the 1240-h Russian spoken corpus (Karpov et al., 2021), we train a deep learning model to derive posterior probabilities for phonological features, such as [ + high, −back], tied to palatalization (Padgett et al., 2024). This enables a detailed analysis of gradient patterns across consonant types and vocalic contexts. Our findings are expected to reveal a continuum of palatalization influenced by phonetic factors, supporting a gradient contrast utilization perspective (Parker, 2015). This study underscores the value of probabilistic approaches in capturing phonetic gradience, bridging computational methods and phonological theory, and advancing understanding of phonetic patterns in Russian and beyond.