Identifying underlying amyloid-beta ($A\beta$) pathology, a hallmark of Alzheimer's disease (AD), typically requires invasive or costly procedures. In addition, differential diagnosis between Alzheimer’s disease (AD) phenotypes and Primary Progressive Aphasia (PPA) variants remains a significant clinical challenge due to overlapping symptomatic profiles. This work evaluates the utility of a deep learning-based spontaneous speech analysis for biomarker prediction in AD and diagnostic classification in PPA. While existing models are biased toward English-language datasets, we benchmark our framework on the IS2021 ADReSSo challenge and extend it to a new Spanish-speaking cohort from the Sant Pau Initiative on Neurodegeneration (SPIN). Our architecture leverages a deep learning system that integrates state-of-the-art encoders for AD detection and audio processing. Our approach integrates acoustic and linguistic features for the automated clinical phenotyping of neurodegenerative diseases.