C-Conformer: Channel-Augmented Conformer for Sound Event Localization and Detection

Three-dimensional sound event localization and detection (SELD) jointly performs event classification, direction-of-arrival (DOA) estimation, and source distance estimation, enabling comprehensive acoustic scene understanding. Recently, the ResNet–Conformer architectures have shown strong performance on SELD. However, the Conformer is primarily designed for single-channel audio recognition and models only temporal dependencies, overlooking inter-channel features for multi-channel spatial perception. To address this, we propose the Channel-augmented Conformer (C-Conformer), which extends the Conformer by explicitly modeling both temporal and inter-channel relationships. C-Conformer preserves temporal self-attention (TSA) and introduces a lightweight channel self-attention (CSA) module using depthwise convolutions to maintain temporal order. The TSA and CSA outputs are fused via learnable weights. To mitigate the distribution gap between synthetic and real data, we further design a Dual-of-Experts (DOE) CSA mechanism, enhancing cross-domain robustness. Extensive experiments show that ResNet–C-Conformer (RCC) consistently outperforms most existing models, achieving up to 8.4% F-score improvement. With DOE CSA, RCC effectively leverages both synthetic and real datasets, surpassing our previous system, which ranked second in the DCASE2025 SELD Challenge.