Intoxicated speech detection: A fusion framework with speaker-normalized hierarchical functionals and GMM supervectors

► We combine two statistic/hierarchical functional subsystems (differing in their speaker normalization method) with three GMM subsystems. ► We improved accuracy after taking into account that the GMM subsystems happened to be capturing lexical information that was distinct between classes only on the training and development sets. ► We show that background speaker normalization is a useful method for speaker state applications. ► Additionally, we show that longer utterances and utterances requiring higher cognitive load are easier to detect.