Order-aware Pairwise Intoxication Detection

Abstract

Alcoholic intoxication has always been and still is known as one of the major causes leading to traffic accidents and in-car conflicts. A system of intoxication detection is established to detect whether a person is intoxicated through the means of machine learning. The system would be able to provide significant assistance in the enforcement of traffic laws, which would ultimately save lives. However, most of the existing systems mainly attach great importance to the tested speaker’s characteristics of current speech, and ignore the existence of personalized differences in speech. To deal with this problem, we focus on modeling the measurable acousic change between the current state and the sober state of a speaker, instead of the current state in the existing scheme only. Furthermore, we are inspired by our discovery that the order-related cues (e.g. gender, time, location) on speaker and trip is largely relevant to alcoholic intoxication. Therefore, we incorporate order-related cues into the speechbased system in order to obtain better performance. Finally, it is demonstrated by extensive experimental results on DiDi Drunk Dataset in real scene that our proposed system achieved a significant improvement from 74.1% to 84.9% in terms of AUC.