Chronic pain and presbyphonia are degenerative conditions that affect most older people, for which there are currently no objective, non-invasive assessment methods. The present pilot study proposes a multitask learning model that is capable of analysing vocal biomarkers (128 acoustic features, including MFCCs, jitter and shimmer) across three tasks simultaneously: pain classification (four levels), presbyphonia detection (binary) and age estimation (regression). For the feasibility analysis, 233 audio recordings describing pain on perception-based scales were used. However, the architecture of the shared layer model comprising 256–128 neurons and 91,143 parameters produced unsatisfactory results. The model demonstrated a perfect accuracy rate in the detection of presbyphonia (n = 26 in the validation set), however, its performance in pain classification was less successful, achieving a mere 23.1% accuracy. The model was able to detect cases that were categorised as ‘Strong’ (recall: 85.7%). While simultaneous multitasking enhanced computational efficiency in comparison with individual models, constraints in multi-class generalisation and a critical reliance on data balancing indicate that, in order to guarantee the clinical applicability of the findings of this preliminary work, it is essential to consider the following: The primary element that must be given due consideration at the outset is the expansion of the dataset. The second element is the incorporation of hybrid architectures (CNN-Transformer), and the third element is validation in real-world settings. Preliminary results support the feasibility of multimodal voice analysis for initial triage; however, improvements are required for reliable diagnostic use in geriatric digital health.