Abstract: This study explored the effect of a speaker’s attachment style (secure, detached, preoccupied, and fearful) on acoustic characteristics of emotional speech. Grammatical but meaningless pseudo sentences were designed, and participants with romantic relationship experiences were recruited. After activating their attachment systems with the subliminal lexical priming paradigm, the participants watched the videos that tended to evoke one of the four basic emotions (happiness, anger, sadness, and fear). Following this, they expressed these sentences with the corresponding emotions to their imagined romantic partners. Based on 14 acoustic parameters per utterance selected using the recursive feature elimination algorithm, Semi-Parametric Repeated Measures Multivariate Analysis of Variance shows significant main effects of attachment style and emotion type, but does not show a significant interaction effect between them. Agglomerative Hierarchical Cluster Analysis shows that, in the acoustic space, “dismissing” and “preoccupied” are the closest, while “secure” is the farthest from other attachment styles. Supervised classification algorithms effectively differentiate the four attachment styles based on 14 acoustic parameters, with prosodic parameters contributing more in terms of feature importance analysis. Furthermore, Accumulated-Local Profiles analysis indicates that the variations in fundamental frequency characteristics among the four attachment styles remain basically consistent across the four emotions, but the differences in timbre and voice quality characteristics are influenced by emotion type. In summary, this study unveils the impact of attachment style on emotional speech and confirms the variations in emotion regulation strategy among individuals with four attachment styles. This provides a scientific foundation for the development of personalized human-machine speech communication technologies.