To determine whether automated lexical dysfluency analysis can differentiate non-fluent variant primary progressive aphasia (nfvPPA) from logopenic variant PPA (lvPPA), which are often difficult to distinguish in early disease stages.

NfvPPA and lvPPA present with overlapping language impairments but distinct underlying mechanisms: nfvPPA shows motor-speech disruptions, while lvPPA involves phonological errors. Automated speech recognition (ASR) systems can aid objective speech analysis but often miss dysfluencies as they prioritize fluent transcription. We developed a forced-alignment–based approach, the Scalable Speech Dysfluency Modeling Lightweight system (SSDM-L), to capture phoneme- and word-level disruptions overlooked by conventional ASR.

Participants included 40 individuals with nfvPPA, 40 with lvPPA, and 27 healthy controls who read aloud the ‘Grandfather passage’. Eight dysfluency variables were extracted using SSDM-L, including insertions, replacements, deletions at both phoneme- and word-levels, and phoneme-level prolongations and repetitions. Group differences were assessed via ANCOVAs controlling for age, education, and disease severity (MMSE, CDR sum-of-boxes). To test clinical validation, we performed correlation analyses with the gold-standard expert Motor Speech Exam (MSE) ratings within the nfvPPA group. Classification performance was assessed by training XGBoost machine-learning models including 5-fold cross-validation.

All features distinguished PPA from controls (p<.001–.004). NfvPPA individuals made more errors than lvPPA individuals on each of the eight features (p<.001–.023). Each feature showed a moderate positive correlation with the combined MSE apraxia/dysarthria score (r = .31–.56; p<.001-.053). Together, the eight features were able to classify nfvPPA vs lvPPA at AUC=.792 [95% confidence interval: .600-.983], and adding age, education, and disease severity improved model performance to AUC=.917 [.805-1.00].

Automated phoneme- and word-level dysfluency analysis accurately distinguishes PPA variants using a brief reading task. This objective, scalable method reduces reliance on expert perceptual judgment and addresses current limitations of ASR, offering a clinically practical tool for differential diagnosis in language-based dementias.