AI-assisted vocal emotion analysis in forensic interview with children: an exploratory study

Remote forensic interviews may restrict access to children's nonverbal cues, potentially constraining affective assessment. This exploratory study examined whether children's emotional states during forensic interviews could be identified using an artificial intelligence (AI)-assisted system that analyzes vocal emotional biomarkers, and whether AI-derived affective indices differ between an AI-assisted interview condition featuring real-time monitoring of children's emotion and a traditional face-to-face interview condition. Fifty-nine children aged 4-8 years participated in simulated forensic interviews, yielding 2,084 utterances. Acoustic features were analyzed post hoc on recorded interview audio using a pretrained speech emotion model to estimate probabilities for happiness, anger, sadness, and neutral emotion. Regression analyses with participant-level clustering were conducted to account for the repeated-measures structure. Results indicated that anger probabilities, as well as anger-to-sadness and neutral-to-sadness ratios, were significantly higher in the AI-assisted condition. However, overall emotional distributions did not indicate increased distress associated with the AI-assisted modality, and dominant happiness did not differ significantly between groups. These findings suggest that AI-based vocal affect analysis may serve as a supplementary observational tool in forensic interviews while supporting the emotional validity of AI-assisted interview conditions. Particularly in contexts where visual cues are limited, AI-assisted approaches may offer a structured means of monitoring children's affective changes without replacing professional judgment.