CHAPTER Purohit_SPRINGERNATURE_2025/IDIAP On Detection of Depression in Parkinson's Disease Patients' Speech: Handcrafted Features vs. Speech Foundation Models Purohit, Tilak Ruvolo, Barbara Orozco-Arroyave, Juan Rafael Magimai-Doss, Mathew Llorente, Juan Ignacio Godino Ed. depression detection Interpretable features Parkinson’s disease Speech Foundation Models EXTERNAL http://publications.idiap.ch/attachments/papers/2025/Purohit_SPRINGERNATURE_2025.pdf PUBLIC Automatic Assessment of Parkinsonian Speech Communications in Computer and Information Science (CCIS) 2646 2025 Springer Nature Switzerland AG Cambridge, MA, USA https://link.springer.com/book/9783032070821 URL This study investigates speech-based depression detection in individuals with Parkinson’s disease (PD), comparing two feature representation approaches: interpretable handcrafted acoustic features and non-interpretable representations derived from Speech Foundation Models (SFMs). The study utilizes the DAIC-WOZ corpus for typical depression (non-PD speech) and the PD-Depression corpus for depression in atypical speech (PD speech). We first examine the viability of handcrafted features and then analyse how acoustic descriptors differ across these conditions. We then evaluate SFM-based representations on the PD-Depression dataset using a cross-validation-based layer selection methodology. Results suggest that handcrafted features yield better classification performance for depression in PD speech compared to SFM embeddings. Analysis shows that while typical depression is marked by pitch instability and reduced vocal quality, speech from PD patients with depression exhibits broader spectral variability, likely due to motor impairments associated with hypokinetic dysarthria.