Today’s overburdened health systems face numerous challenges, exacerbated by an aging population. Speech emerges as a ubiquitous biomarker with strong potential for the development of low-cost, remote testing tools for several diseases. In fact, speech encodes information about a plethora of diseases, which go beyond the so-called speech and language disorders, and include neurodegenerative, psychiatric, and respiratory diseases.
Recent advances in speech processing and machine learning have enabled the automatic detection of these diseases. Despite promising results, this research area faces challenges, primarily due to dataset limitations and the overlap of speech-affecting diseases, which often coexist and produce similar speech manifestations.
These challenges guide our latest research, where we discuss the characterization of normative speech. Similar to common blood tests, we explore reference intervals for interpretable speech features (acoustic and linguistic) as a first step toward adopting speech analysis for multidisease screening. We leverage deviations from these references to detect Alzheimer’s and Parkinson’s diseases using different classifiers, namely Neural Additive Models for enhanced interpretability.
Additionally, we explore bridging black-box models and interpretability by using large language models to annotate high-level, low-dimensional, interpretable characteristics of speech transcriptions, termed macro-descriptors—such as text coherence and lexical diversity. Using only four macro-descriptors, we outperformed conventional text-based Alzheimer’s disease detection methods.
