I'm Fine, But My Voice Isn't: Cross-Modal Affective Dissonance Detection for Reflective Journaling

Original Abstract

Digital journaling creates an authenticity gap: users consciously translate raw emotions into text, often sanitizing narratives even in private writing. We formalize this as Cross-Modal Affective Dissonance Detection (CADD), a directional three-way classification distinguishing Masking (positive text, negative acoustics), Coping (negative text, positive acoustics), and Congruent utterances, grounded in Gross's process model of emotion regulation. We present three further contributions: (i) CADD-Journal, a 1,800-sample TTS dataset with a shared-sentence-pool design that provably isolates acoustic signal from textual content; (ii) DACM, a dual-encoder model with asymmetric cross-modal attention that re-solves a gradient degeneracy in pooled fusion, achieving macro-F1 0.711 - with a four-step ablation demonstrating that asymmetric attention is the dominant driver (+ 0.242) while the DIM is effective only on cross-modal features (+0.033); and (iii) a domain gap quantification: zero-shot evaluation across three naturalistic corpora reveals a substantial gap between TTS-trained models and real speech, and we identify two concrete requirements for future in-the-wild corpus construction. ReflectJournal, a proof-of-concept iOS application, operationalizes the framework and provides a deployment platform for naturalistic data collection.