Acoustic sleepiness detection: framework and validation of a speech-adapted pattern recognition approach

Area of Science:

• Speech analysis
• Pattern recognition
• Sleep science

Background:

• Sleepiness detection is crucial for safety and performance.
• Current methods often require intrusive sensors or calibration.
• Automatic, real-time detection offers a non-obtrusive alternative.

Purpose of the Study:

• To develop a general framework for detecting sleepiness states using speech characteristics.
• To evaluate the effectiveness of an automatic, real-time speech analysis approach for sleepiness detection.

Main Methods:

• A framework combining prosody, articulation, and speech-quality features was developed.
• A total of 45,088 features were extracted per speech sample.
• A support-vector machine (SVM) classifier was trained after feature selection and dimensionality reduction.

Main Results:

• The best model, an SVM, achieved 86.1% accuracy in classifying sleepiness.
• The study involved a sleep deprivation experiment with 12 participants.
• The approach demonstrated the feasibility of non-intrusive, real-time sleepiness detection.

Conclusions:

• Speech analysis provides a viable, non-obtrusive method for real-time sleepiness detection.
• The proposed framework effectively utilizes acoustic features for accurate sleepiness classification.
• This technology has potential applications in monitoring driver fatigue and other safety-critical domains.