The present study investigates how variation in acoustic measures and lexical semantic properties affect the comprehension of threatening speech. Two experiments were conducted in British English, where participants had to rate the threat of three types of sentences: sentences neutral in prosody and semantics, sentences containing only prosodic threat, and sentences containing only semantic threat. They also rated the sentences for arousal and valence, and categorised them by type of emotional expression. Threat ratings were analysed via Bayesian ordered-logistic regressions using acoustic measures and affective norms as regressors. Classification was assessed via simple Bayesian categorical (Dirichlet prior) models. Acoustic properties of stimuli were compared through a Bayesian method for comparison of means to test whether neutral stimuli differ, on average, from their threatening counterparts. Acoustic but not semantic properties affected ratings of threatening prosody (increased pitch and decreased voice quality predicting increased threat), while the reverse was true for ratings of threatening semantics (increased arousal and decreased valence predicting increased threat). Using data from the first speaker (first experiment) the models were able to predict ratings given to sentences produced by the second speaker (second experiment). Furthermore, threatening sentences of both types tended to be categorized as angry or enraged. These results are interpreted as supporting the sufficiency of the selected measures (pitch, voice quality, arousal, valence) to characterise threat. Although multidimensional models may be better suited for describing threatening prosody/semantics, the present approach demonstrates that selected bidimensional features are sufficient to identify and predict threat comprehension efficiently and accurately.