Automatic Speech Recognition systems errors for objective sleepiness detection through voice
(3 minutes introduction)
| Vincent P. Martin (LaBRI (UMR 5800), France), Jean-Luc Rouas (LaBRI (UMR 5800), France), Florian Boyer (LaBRI (UMR 5800), France), Pierre Philip (SANPSY (USR 3413), France) |
|---|
Chronic sleepiness, and specifically Excessive Daytime Sleepiness (EDS), impacts everyday life and increases the risks of accidents. Compared with traditional measures (EEG), the detection of objective EDS through voice benefits from its ease to be implemented in ecological conditions and to be sober in terms of data processing and costs. Contrary to previous works focusing on short-term sleepiness estimation, this study focuses on long-term sleepiness detection through voice. Using the Multiple Sleep Latency Test corpus, this study introduces new features based on Automatic Speech Recognition systems errors, in an attempt to replace hand-labeled reading mistakes features. We also introduce a selection feature pipeline inspired by clinical validation practices allowing ASR features to perform on par with the state-of-the-art systems on short-term sleepiness detection through voice (73.2% of UAR). Moreover, we give insights on the decision process during classification and the specificity of the system regarding the threshold delimiting the two sleepiness classes, Sleepy and Non-Sleepy.
