Kashaf Gulzar, Korbinian Riedhammer, Elmar Nöth, Andreas K. Maier, Paula Andrea Pérez-Toro · Speech Communication 2026 · 2026
DOI: 10.1016/j.specom.2026.103497
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Speech-based detection of cognitive impairment (CI) offers a promising non-invasive approach for early diagnosis, yet performance disparities across demographic and clinical subgroups remain underexplored, raising concerns around fairness. This study evaluates acoustic-based classification of CI and comorbid depression on the DementiaBank Pitt Corpus, with a systematic bias analysis for CI detection. We compare traditional features (MFCCs, eGeMAPS) with contextualized speech embeddings from Wav2Vec 2.0 (W2V2). For CI detection, higher-layer W2V2 embeddings outperform traditional features (UAR up to 80.6%), whereas comorbid depression classification yields lower performance, with modest improvements from low- and mid-level W2V2 layers. Cross-task generalization is limited, suggesting that each task relies on distinct acoustic-semantic cues. We evaluate downstream performance across gender, age, and depression subgroups and complement it with subgroup-level AUCs to distinguish threshold bias from predictive bias. Analysis reveals that balancing mitigates some threshold bias, but predictive bias persists in W2V2, with lower separability for female and younger speakers (AUC gaps up to 0.055 and 0.098, respectively). Depression disparities are the strongest, with near-ceiling separability; however, they are limited by the small size of the depressed subgroup. These findings show that data balancing alone does not ensure equitable performance, underscoring the need for fairness-aware evaluation in clinical speech applications.
Last synced
This work has 2 recorded citations, but citing papers have not been linked locally yet.
No comments yet — start the discussion below.