Public defence: Syed Zohaib Hassan

The task of investigative interviewing allegedly abused children is demanding and complex. Child Protection Services (CPS) and law enforcement professionals must be skilled while employing appropriate questioning techniques, as errors can lead to case dismissals due to inadmissible evidence. Although evidence-based practices have been developed to conduct these interviews effectively, current interview training programs face significant limitations in terms of cost-effectiveness, scalability and efficacy, and lack sufficient practice opportunities. 

We have developed and evaluated an AI-powered child avatar system for investigative interview training to address these challenges. The system combines dialogue models, various interactive environments, and speech technologies to provide an immersive training environment. The system also integrates real-time feedback mechanisms to create realistic, interactive training scenarios that enhance interviewer skill development while addressing the practical limitations of traditional approaches.

A series of Quality of Experience (QoE) user studies with CPS professionals evaluated the system across multiple dimensions. Our studies revealed high ratings for realism and presence for Virtual Reality (VR)-based avatars, with participants reporting strong potential for skill development. The dialogue evaluation identified strengths in simulating realistic child-like communication while highlighting technical challenges with responsiveness and emotional depth. 

The further exploration involved the comparative analysis of different interactive environments (VR, 2D monitor, audio-only, and text-based) that demonstrated VR delivered significantly higher engagement and presence ratings; while all other modalities maintained comparable effectiveness for skill development.

Another contribution of this work is developing and validating a Large Language Model (LLM) based feedback mechanism that demonstrated substantial improvement in questioning techniques among professionals. Participants receiving direct feedback showed a statistically significant increase in best practice questioning that persisted over time. Our research also addressed components of emotional realism, finding that a four-emotion framework achieved optimal reliability for classifying emotional patterns in child interviews. While fine-tuning LLM for incorporating disfluencies in the speech showed statistical significance in enhancing the perceived spontaneity of the generated speech. A systematic investigation of LLMs' capability to generate age-appropriate responses demonstrated that while LLMs can approximate child-like language with reasonable accuracy for younger children, they struggle with older children's more nuanced linguistic patterns. 

This thesis contributes to both the technical and training effectiveness aspects of investigative interview training. By addressing fundamental challenges in system development, user experience, feedback mechanisms, and emotional realism, this research establishes a foundation for more effective training solutions that can meaningfully improve the quality of investigative interviews with children, ultimately supporting better outcomes for maltreated children.