Phonetisch-akustische Schläfrigkeitsdetektion.

This paper describes the development and validation of a phonetic-acoustic measurement procedure for a speech-based detection of sleepiness. The advantages of this automatic real-time approach are that obtaining speech data is unobtrusive and free from sensor application and calibration efforts. The chosen measurement process follows the speech-adapted steps of pattern recognition: (1) recording speech, (2) computation of 170 features describing prosody, articulation, and voice quality, (3) machine learning, and (4) evaluation. In a sleep deprivation study, a total of 380 simulated driver assistance samples ( n=32; 8:00 p.m.-4:00 a.m.) were recorded. One self and two observer assessments were used to obtain a Karolinska Sleepiness Scale (KSS) value, which served as an external validation reference. Features that proved to be especially sensitive to sleepiness are cepstral coefficients, formant bandwidth, intensity, and spectral measures. The best machine learning method, the support vector machine (SVM), achieved a significant validation correlation of r=0.46 in predicting sleepiness on unseen speakers. [ABSTRACT FROM AUTHOR]