Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting Mental Manipulation in Speech via Synthetic Multi-Speaker Dialogue

Run Chen, Wen Liang, Ziwei Gong, Lin Ai, Julia Hirschberg

TL;DR

This work addresses detecting mental manipulation in spoken dialogue by introducing SpeechMentalManip, a synthetic, voice-consistent audio benchmark derived from a text-based manipulation dataset. It uses a two-phase TTS pipeline to generate multi-speaker conversations and evaluates few-shot audio-language models alongside human annotators to compare text and speech modalities. Findings show higher precision but notably lower recall in audio-based detection, accompanied by greater annotation ambiguity and lower inter-annotator agreement, highlighting modality-driven challenges. The study emphasizes modality-aware evaluation and safety considerations for multimodal dialogue systems and outlines directions for expanding voice diversity and improving audio-first annotation practices.

Abstract

Mental manipulation, the strategic use of language to covertly influence or exploit others, is a newly emerging task in computational social reasoning. Prior work has focused exclusively on textual conversations, overlooking how manipulative tactics manifest in speech. We present the first study of mental manipulation detection in spoken dialogues, introducing a synthetic multi-speaker benchmark SPEECHMENTALMANIP that augments a text-based dataset with high-quality, voice-consistent Text-to-Speech rendered audio. Using few-shot large audio-language models and human annotation, we evaluate how modality affects detection accuracy and perception. Our results reveal that models exhibit high specificity but markedly lower recall on speech compared to text, suggesting sensitivity to missing acoustic or prosodic cues in training. Human raters show similar uncertainty in the audio setting, underscoring the inherent ambiguity of manipulative speech. Together, these findings highlight the need for modality-aware evaluation and safety alignment in multimodal dialogue systems.

Detecting Mental Manipulation in Speech via Synthetic Multi-Speaker Dialogue

TL;DR

This work addresses detecting mental manipulation in spoken dialogue by introducing SpeechMentalManip, a synthetic, voice-consistent audio benchmark derived from a text-based manipulation dataset. It uses a two-phase TTS pipeline to generate multi-speaker conversations and evaluates few-shot audio-language models alongside human annotators to compare text and speech modalities. Findings show higher precision but notably lower recall in audio-based detection, accompanied by greater annotation ambiguity and lower inter-annotator agreement, highlighting modality-driven challenges. The study emphasizes modality-aware evaluation and safety considerations for multimodal dialogue systems and outlines directions for expanding voice diversity and improving audio-first annotation practices.

Abstract

Mental manipulation, the strategic use of language to covertly influence or exploit others, is a newly emerging task in computational social reasoning. Prior work has focused exclusively on textual conversations, overlooking how manipulative tactics manifest in speech. We present the first study of mental manipulation detection in spoken dialogues, introducing a synthetic multi-speaker benchmark SPEECHMENTALMANIP that augments a text-based dataset with high-quality, voice-consistent Text-to-Speech rendered audio. Using few-shot large audio-language models and human annotation, we evaluate how modality affects detection accuracy and perception. Our results reveal that models exhibit high specificity but markedly lower recall on speech compared to text, suggesting sensitivity to missing acoustic or prosodic cues in training. Human raters show similar uncertainty in the audio setting, underscoring the inherent ambiguity of manipulative speech. Together, these findings highlight the need for modality-aware evaluation and safety alignment in multimodal dialogue systems.
Paper Structure (22 sections, 1 equation, 5 figures, 6 tables)

This paper contains 22 sections, 1 equation, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Mental Manipulation in Dialogue
  4. LMMs safety
  5. Synthetic Speech
  6. Method
  7. Dataset and Voice Pool
  8. Multi-Speaker TTS Audio Generation
  9. Phase 1: Turn-level audio generation.
  10. Phase 2: Conversation reconstruction.
  11. Model Selection
  12. Experiment Setup
  13. Few-shot Detection Pipeline
  14. Evaluation Protocol and Metrics
  15. Audio-only Few-shot Detection Results
  16. ...and 7 more sections

Figures (5)

  • Figure 1: An example dialogue from the SpeechMentalManip dataset. The Qwen2.5 model is given the audio (transcript shown for clarity), but fails to detect manipulation.
  • Figure 2: Two-phase pipeline for TTS audio generation and conversational reconstruction.
  • Figure 3: Pair-wise Cohen's Kappa between Human Annotators for Text modality
  • Figure 4: Pair-wise Cohen's Kappa between Human Annotators for Audio modality
  • Figure 5: Annotation interface. Annotators first reviewed task instructions and the definition of mental manipulation (Guideline tab), then labeled the same dialogue under text-only and audio-only conditions in separate tabs.