Summary: A new deep learning algorithm may help children affected by abuse disclose information about their experiences.

Source: USC

Can Artificial Intelligence Improve Forensic Interviews with Child Abuse Victims?

When children are victims of crimes, their legal testimony often comes from structured forensic interviews. Because many child victims are traumatized or abused by caregivers, they may be reluctant to disclose what happened. Interviewers follow strict protocols designed to gather accurate, relevant information while minimizing retraumatization and avoiding leading or coercive questions.

Researchers at the University of Southern California (USC) investigated whether artificial intelligence (AI) and deep learning techniques can support forensic interviewers by analyzing speech and interaction patterns to identify the approaches that yield the most productive responses from children. Their work, presented at the 2018 ACM International Conference on Multimodal Interaction in Boulder, Colorado, explores how computational tools might supplement training and real-time interviewing decisions without replacing professional judgment.

The paper documents a multidisciplinary collaboration between USC Viterbi School of Engineering’s Signal Analysis and Interpretation Laboratory (SAIL) and the USC Gould School of Law. Lead contributors include doctoral students Victor Ardulov and Manojkumar Prabakaran Abitha and SAIL founder Shri Narayanan, working with child witness expert Professor Thomas D. Lyon and his team. Together they developed methods to automatically detect and categorize linguistic and paralinguistic features—such as emotional tone, pitch, pauses, and turn-taking—that appear to influence how much information children share during forensic interviews.

For interviewers, the central challenge is asking the right questions, at the right moment, and in the right way to encourage truthful, detailed disclosures without introducing bias. Prior psychological and legal research has shown that rapport building, question phrasing, tone, pacing, and even the order of questions all affect a child’s willingness and ability to provide accurate accounts. This USC study represents one of the first attempts to apply custom speech- and language-analysis software to quantify those subtleties automatically and to search for patterns tied to interviewing effectiveness.

Method and Data

The team analyzed anonymized audio recordings and transcripts of two hundred forensic interviews collected by Professor Lyon from child abuse investigations. Transcripts were coded across multiple dimensions and then processed with models developed at SAIL. These computational models combine speech processing, speaker diarization (measuring who speaks and for how long), emotional and prosodic analysis (changes in pitch, loudness, and rhythm), and interaction metrics (such as pause length and mirroring of speech pace between interviewer and child).

Custom machine learning models were trained to detect patterns of interaction and to relate specific interviewer behaviors to measures of interview productivity—the amount and quality of relevant information elicited without leading the child. Rather than replacing human expertise, these tools were intended to augment researchers’ ability to spot subtle patterns that are difficult and time-consuming to count manually.

Key Findings

The findings largely align with established results in legal and developmental psychology but add new computational evidence about how linguistic and vocal features influence disclosure. Forensic interviews typically proceed in two phases: a rapport-building phase that focuses on neutral conversation to establish comfort and trust, and an information-gathering phase that addresses the alleged abuse.

In the study, children’s responses varied considerably with age. Younger children were more affected by the emotional content and phrasing of the interviewer’s words during rapport-building and information-gathering stages. Emotional tone, reassurance, and age-appropriate language related to greater openness among younger participants. Older children, by contrast, showed greater sensitivity to paralinguistic features—how questions were vocalized—such as pitch, loudness, and pacing. When interviewers matched or complemented the child’s speech pace, the interaction tended to be more productive.

Other measurable factors included the length and placement of pauses, the amount of time allotted to the child to respond, and the interviewer’s tendency to repeat or reformulate questions. The automated analyses highlighted sequences of turns and question types that correlated with richer, more detailed accounts without apparent coercion.

Potential Applications and Next Steps

The research suggests several practical directions for AI-assisted tools that could enhance forensic interviewing. One clear application is training: simulated interviews or virtual assistants could provide feedback to new or experienced interviewers on question phrasing, timing, and tone based on patterns learned from large datasets. Another possibility is a real-time aide—an automated transcription and suggestion system that highlights words, phrases, or interaction patterns that may warrant follow-up questions or indicate areas of inconsistency.

Building reliable, ethical tools depends on large, well-annotated datasets of child–interviewer interactions and on rigorous models that account for developmental and contextual differences. The researchers compare this approach to autocomplete systems that predict likely continuations from massive historical input, but emphasize that any deployed system must be carefully validated and used as an aid rather than a determiner of truth. Future research will focus on more sophisticated sequence models that identify which specific question sequences and interactions consistently produce accurate, nonleading disclosures across ages and contexts.

About this research article

Source: Amy Blumenthal, USC

Publisher: Organized by NeuroscienceNews.com

Image source: Image in the public domain

Original research: Presented at the 2018 ACM International Conference on Multimodal Interaction

This multidisciplinary work demonstrates how speech analysis, emotion detection, and interaction modeling can provide new quantitative insight into forensic interviewing practices with child abuse victims. Carefully developed and validated AI tools have the potential to enhance interviewer training, support ethical and effective questioning strategies, and ultimately improve the quality of information gathered from vulnerable child witnesses.