Sleep deprivation detected by voice analysis.

Area of Science:

• Neuroscience
• Acoustics
• Machine Learning

Background:

• Sleep deprivation significantly impacts individuals and society.
• Current methods for detecting sleep deprivation lack speed and objectivity.
• Objective vocal biomarkers for sleep deprivation are needed.

Purpose of the Study:

• To develop and validate an objective method for detecting sleep deprivation using automated voice analysis.
• To enhance the interpretability of machine learning models for sleep deprivation detection.
• To identify specific vocal changes associated with sleep deprivation.

Main Methods:

• Utilized automated acoustic analyses of voice recordings.
• Employed a generic auditory representation based on spectro-temporal modulation analysis.
• Applied reverse correlation for model interpretability.
• Compared performance against state-of-the-art speech features.

Main Results:

• Generic auditory features successfully detected sleep deprivation with accuracy comparable to existing methods.
• Identified two distinct vocal effects: changes in prosody (slow temporal modulations) and voice quality (spectral features).
• Observed significant individual variability in the balance of these vocal effects.

Conclusions:

• Automated voice analysis provides a viable, objective method for detecting sleep deprivation.
• Vocal biomarkers reflect both explicit (subjective sleepiness) and implicit consequences of sleep deprivation.
• Individualized vocal markers of sleep deprivation warrant further investigation for clinical applications.