How to Design a Relevant Corpus for Sleepiness Detection Through Voice?

Area of Science:

• Speech analysis
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Voice processing is utilized for diagnosing various neurological and psychological conditions, including Parkinson's disease, Alzheimer's disease, and depression.
• Despite advancements, automatic detection of sleepiness using voice analysis has not achieved sufficient performance for clinical applications, even with deep learning.

Purpose of the Study:

• To investigate the hypothesis that limitations in current automatic sleepiness detection performance stem from the design of existing speech corpora.
• To refine the understanding of 'sleepiness' in the context of ground-truth labels for voice analysis research.
• To propose guidelines for creating improved speech corpora for sleepiness detection.

Main Methods:

• Conceptual refinement of 'sleepiness' and its relation to ground-truth labeling in speech analysis.
• In-depth analysis of four speech corpora, identifying methodological choices and potential biases.
• Development of recommendations for future corpus design in sleepiness detection research.

Main Results:

• Identified specific methodological choices and inherent biases within existing speech corpora that may impede accurate sleepiness detection.
• Highlighted the critical role of corpus design in the success or failure of automatic speech analysis for detecting specific conditions like sleepiness.

Conclusions:

• The design and inherent biases of speech corpora are significant factors limiting the performance of automatic sleepiness detection systems.
• Establishing clear definitions and robust methodologies for corpus creation is essential for advancing sleepiness detection through voice analysis.