ACOUSTIC SLEEPINESS ANALYSIS-AN ACOUSTICAL PATTERN RECOGNITION APPROACH FOR DETECTING SLEEPINESS FROM RADIO COMMUNICATION

摘要

A promising approach for the real-time detection of sleepiness of Air Traffic Controller (ATC) and pilots is the acoustic sleepiness analysis. Sleepiness related changes in speech refer to three speech feature classes (a) ,,prosody" (flat intonation, low pitch, unexpected breaks in rhythm, slowed speech rate), (b) ,,articulation" (slurred speech) und ,,voice quality" (softening, lenisation). Method. Within-subject sleep deprivation design; 20.00h to 04.00h; N=23; every hour Standford Sleepiness Scale and speech recordings. Acoustic Analysis. The procedure of acoustic sleepiness analysis includes the following steps: (1) recording of speech corpus (22,1 kHz; 16 bit; mono; spontaneous speech, manual segmentation of the german word "mit" [with], 72 speech samples), (2) extracting speech features set (160 raw and 160 z-normalized features; prosody, articulation and speech quality features (e.g. Jitter, HNR, MFCCs), (3) feature selection and (4) using classification algorithm (MLP, kNN, SVM). Results. The best model (forward selection, 1-NN) offers a mean accuracy rate of 95.0% for males and 94.3% for females for the two-class problem of recognizing sleepy vs. alert speech. Discussion. Critical sleepiness states can be identified semi-automatically with high accuracy out of a single one-syllable word. Our results are limited mainly by the fact that the findings were based on noise-free recordings. Further studies should try to employ real VHF analogue radio communication for the sleepiness prediction of ATCs and pilots.