<p>The objective of Emotion Recognition in Conversation (ERC) is to analyze emotional states within a conversational context. Existing methods often suffer from limited long-range dependency modeling and insufficient multi-scale feature capture, where emotional cues are diluted during spatial propagation and discriminative high-frequency components are suppressed by over-smoothed spectral representations. To address this challenge, we propose a novel Dual-Branch Spatio-Spectral Fusion (DBSF) approach from a graph spectral perspective. To prevent contextual dilution, the spatial branch features an enhanced Multi-hop Feature Propagation (MFP) module driven by Kolmogorov-Arnold Networks (KANs) to capture the dynamic emotion flow. By utilizing spline-enhanced attention, this module captures complex speaker interactions across expanded receptive fields, thereby alleviating the limitations of linear projections. To mitigate feature homogenization and the resulting over-smoothing problem, the spectral branch employs Chebyshev Wavelet Approximation for multi-scale analysis. It utilizes scaling functions to capture slowly-varying global moods, while band-pass wavelets explicitly isolate and preserve fleeting localized emotional shifts that are typically suppressed in spatial aggregation. By jointly modeling emotion dynamics in the spatial and spectral domains, the proposed framework helps alleviate over-smoothing and preserves discriminative features in long-range contexts. Experiments on MELD and IEMOCAP demonstrate that our approach performs competitively against existing baselines, preserving fine-grained discriminative features across extended conversational contexts.</p>