Deep learning-derived measures of sound-level accuracy in primary progressive apraxia of speech: A feasibility pipeline with descriptive evidence from two cases

Primary Progressive Apraxia of Speech (PPAOS) is a neurodegenerative motor speech disorder marked by disrupted speech planning and programming. This study demonstrates the implementation and interpretability of phonological-class-based posterior probability values generated by Phonet (a deep learning model) in two individuals with PPAOS (one phonetic, one prosodic predominant), across different speech rates and delayed auditory feedback (DAF) settings. Phonet generated frame-level posterior probabilities for phonological classes, which were aggregated and aligned with segment durations. The prosodic subtype showed higher, more stable articulatory values and improvements under DAF, while the phonetic subtype showed greater variability and sensitivity to condition changes. Results were consistent with expert articulatory judgements of sound errors. These findings demonstrate Phonet's potential to provide interpretable, segment-level measures of articulatory performance and suggest its utility in monitoring treatment-related articulatory changes in motor speech disorders. Further research is needed to assess its broader clinical applicability.