SCoPE: Shift-Aware Speaker Priors Boost Emotion Recognition in Conversations

Emotions in conversations are rarely stable—they shift gradually and are often speaker-specific. Yet most existing emotion recognition models rely heavily on overt cues (facial expressions, tone, words), making them brittle when signals degrade due to occluded faces, slang, or microphone noise. The research team behind SCoPE addresses this gap by explicitly modeling the non-apparent factors that govern emotional transitions. SCoPE is a lightweight module that builds a personalized prior for each speaker based on their emotional history, capturing tendencies like escalation or gradual cooling. It then predicts whether an emotion shift is likely to occur, and fuses the prior with incoming multimodal evidence using a precision-weighted logit integration—a Bayesian-inspired product-of-experts approach.

On the IEMOCAP dataset, a standard benchmark for multimodal ERC, SCoPE achieved superior performance over recent state-of-the-art models, including those with far larger architectures. The key innovation is its shift-aware fusion: when emotions are likely stable, the model leans on historical priors; when a shift is predicted (e.g., after surprising news), it prioritizes fresh multimodal evidence. This makes SCoPE particularly robust to noise and ambiguity, as it doesn't require perfect input signals to infer emotion accurately. The paper is currently under review at Cognitive Computation, but its practical implications are clear—better, more human-like emotion sensing for chatbots, mental health analysis, and customer service systems that must handle messy, real-world conversations.