BEAT2AASIST model with layer fusion for ESDD 2026 Challenge
Abstract
Recent advances in audio generation have increased the risk of realistic environmental sound manipulation, motivating the ESDD 2026 Challenge as the first large-scale benchmark for Environmental Sound Deepfake Detection (ESDD). We propose BEAT2AASIST which extends BEATs-AASIST by splitting BEATs-derived representations along frequency or channel dimension and processing them with dual AASIST branches. To enrich feature representations, we incorporate top-k transformer layer fusion using concatenation, CNN-gated, and SE-gated strategies. In addition, vocoder-based data augmentation is applied to improve robustness against unseen spoofing methods. Experimental results on the official test sets demonstrate that the proposed approach achieves competitive performance across the challenge tracks.
Cite
@article{arxiv.2512.15180,
title = {BEAT2AASIST model with layer fusion for ESDD 2026 Challenge},
author = {Sanghyeok Chung and Eujin Kim and Donggun Kim and Gaeun Heo and Jeongbin You and Nahyun Lee and Sunmook Choi and Soyul Han and Seungsang Oh and Il-Youp Kwak},
journal= {arXiv preprint arXiv:2512.15180},
year = {2025}
}
Comments
3 pages, 1 figure, challenge paper