Datatehokkuus ja robustisuus eri käyttöympäristöissä puhesignaaliin perustuvassa terveydentilan arvioinnissa
Year: 2026
Type: dissertation
Abstract: Speech-based biomarking of health exploits the link between speech production and health to detect and monitor health conditions in a non-invasive and cost-efficient manner, with smartphones enabling broad access. A key challenge is domain robustness under data scarcity: medically labeled speech databases are small, and training–testing mismatches in speaker-level and environmental factors often degrade performance. This dissertation targets these issues in the context of classification tasks, emphasizing data-efficient feature extraction and classification within pipeline systems. Three directions were investigated. First, data-efficient feature extraction was explored using mel-frequency cepstral coefficients (MFCCs) with long frames that de-emphasize vocal-tract detail, glottal-source MFCCs that exclude vocal-tract resonances, and transformer-based self-supervised representations (wav2vec 2.0, HuBERT). Across binary and multiclass settings, these representations improved accuracy over conventional features, indicating reduced task-specific redundancy and better generalization with limited data. Second, data-efficient classification was studied via hierarchical designs that decompose multiclass problems into clinically meaningful subtasks. The hierarchical support vector machine (SVM) model outperformed one-vs-one (OvO) and one-vs-rest (OvR) baselines, possibly due to improved class balance and lower model complexity. Third, domain robustness was assessed explicitly in mismatch evaluations. In cross-database evaluation, self-supervised features reduced the performance drop relative to single-database training. For speech transmitted in telephony channels, a test-time supervised preprocessing model was trained to reconstruct high-quality mel-spectrograms from encoded–decoded signals, recovering performance close to non-mismatch conditions and surpassing a baseline. Overall, the dissertation yields three takeaways. First, feature representations that reduce sensitivity to speaker- and channel-specific variation lead to more accurate classifiers than conventional baselines, with self-supervised transformer embeddings offering the most consistent gains. Second, imposing clinically meaningful structure on multiclass classification improves learning efficiency, outperforming standard multiclass approaches while keeping the decision process simple and better balanced. Third, self-supervised representations offer clear benefits under domain mismatch, reducing cross-database degradation and mitigating codec-related mismatches. Despite these advances, remaining gaps under data scarcity motivate further work toward nonmismatch performance in real-world deployments of speech-based biomarking of health.
Source: Aaltodoc (Aalto University)
Author: Tirronen, Saska
Institution: Hospital Universitario Príncipe de Asturias
Language: English
Topic: Voice and Speech Disorders
Subfield: Speech and Hearing
Field: Health Professions
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7162221447
