Detection of alcohol intoxication using voice features: a controlled laboratory study.
OBJECTIVES Devices such as mobile phones and smart speakers could be useful to remotely identify voice alterations associated with alcohol intoxication, which could be used to deliver just-in-time interventions, but data to support such approaches for the English language are lacking. In this controlled lab study, we compare how well English spectrographic voice features identify alcohol intoxication. METHODS 18 participants (72% male, aged 21-62 y) read a different randomly-assigned tongue twister prior to drinking and each hour for up to 7 hours after drinking a weight-based dose of alcohol. Vocal segments were cleaned and split into 1 second windows. We built support vector machine models for detecting alcohol intoxication, defined as breath alcohol concentration [BrAC] >0.08%, comparing the baseline voice spectrographic signature to each subsequent timepoint and present ensemble examine accuracy with 95% confidence intervals (CIs). RESULTS Alcohol intoxication was predicted with an accuracy of 98% (95% CI 97.1 to 98.6); mean sensitivity = .98; specificity = .97; positive predictive value = .97; and negative predictive value = .98. CONCLUSIONS In this small controlled lab study, voice spectrographic signatures collected from brief recorded English segments were useful in identifying alcohol intoxication. Larger studies using varied voice samples are needed to validate and expand models.