Past Event: Monday Colloquium: Using articulatory-informed neural network models to understand gradient phonetic variation (Kevin Tang, Heinrich Heine University, Düsseldorf)

Talk Abstract: Speech articulation is highly sensitive to cognitive, neurological, and linguistic factors, manifesting as subtle phonetic variation in different conditions such as phonological environments, intoxication, neurological disorders, and second language acquisition. This talk presents an articulatory-informed neural network approach to modeling and quantifying such variations, leveraging posterior probabilities of phonological features as a measure of lenition (Tang et al., 2023). Prior work has demonstrated that these neural network-based metrics align with known lenition patterns in Spanish stops, and that they complement and offer advantages over traditional acoustic measures (Tang et al., 2024; Wayland et al., 2023a). We applied this method to examine second language Spanish learners, showing that lenition patterns develop with proficiency, yet approximant-like realizations remain elusive even in immersive environments (Wayland et al., 2023c, 2024c). 

Given its success in evaluating fine-lenition differences between groups and within individuals, we extended it to speech by atypical populations. First, we analyzed speech changes associated with intoxication in English speakers, revealing systematic deaffrication and spirantization effects (Wayland et al., 2023b, 2023c). Second, we applied it to speech of Spanish-speaking Parkinson’s disease patients (PD), revealing that PD exhibited a higher degree of lenition in their voiceless stops compared to healthy controls and that lenition is more advanced for dental stops (Wayland et al., 2024a). Finally, we evaluated the speech of English-speaking PD and atypical parkinsonism (APD) uncovering distinct lenition patterns, with PD patients exhibiting greater articulatory stability than APD patients, highlighting the potential of lenition as a diagnostic marker (Wayland et al., 2024b). These findings underscore the utility of articulatory-informed neural network models in capturing gradient phonetic variation and provide insights into how motor control, cognitive function, and linguistic experience shape speech production.