Language-Independent Sleepy Speech Detection.

Area of Science:

• Speech processing
• Machine learning
• Cognitive science

Background:

• Prolonged sleepiness impairs cognitive and physical performance, increasing accident risk.
• Speech analysis offers an accessible method for automated sleepiness detection.
• Existing methods lack linguistic independence, limiting broad application.

Purpose of the Study:

• To investigate the feasibility of linguistic-independent sleepiness detection from speech signals.
• To compare the performance of deep learning and machine learning models in detecting sleepiness across languages.

Main Methods:

• Trained ResNet50 (deep learning) and five machine learning models on English speech data.
• Validated models using English speech data.
• Tested model performance on German speech data to assess linguistic independence.

Main Results:

• Deep learning models, particularly ResNet50, significantly outperformed machine learning models.
• ResNet50 achieved high accuracy (0.96), sensitivity (0.92), specificity (0.99), and geometric mean (0.95) on German test data.
• Results demonstrate accurate sleepiness detection irrespective of the language.

Conclusions:

• Sleepiness detection from speech is feasible and accurate, independent of linguistic variations.
• Deep learning offers a robust solution for developing universal sleepiness detection systems.
• This technology has significant potential for preventing performance degradation and accidents.