BEAT2AASIST model with layer fusion for ESDD 2026 Challenge

View on arXiv ← Back to list

Authors: Sanghyeok Chung, Eujin Kim, Donggun Kim, Gaeun Heo, Jeongbin You, Nahyun Lee, Sunmook Choi, Soyul Han, Seungsang Oh, Il-Youp Kwak

Published: 2025-12-17 08:24:12+00:00

AI Summary

The paper introduces BEAT2AASIST, an anti-spoofing model designed for the Environmental Sound Deepfake Detection (ESDD) 2026 Challenge. This approach extends the BEATs-AASIST baseline by implementing dual AASIST branches that process frequency-split or channel-split BEATs features. The system utilizes multi-layer fusion and vocoder-based data augmentation to enhance feature robustness and generalization against unseen spoofing methods.

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.

Key findings

The proposed BEAT2AASIST models achieved competitive performance across both ESDD 2026 challenge tracks. The best ensemble systems secured a Test EER of 1.60% for Track 1 and achieved a strong third-place ranking on Track 2 with a Test EER of 0.35%.

Approach

The BEAT2AASIST architecture uses a BEATs encoder coupled with dual AASIST branches, processing feature representations split either frequency-wise or channel-wise. Feature enrichment is achieved using top-k transformer layer fusion (concatenation, CNN-gated, or SE-gated strategies). Additionally, vocoder-based data augmentation is employed to improve generalization to diverse spoofing patterns.

Datasets

EnvSDD dataset (ESDD 2026 Challenge)

Model(s)

BEAT2AASIST (BEATs encoder, AASIST, dual-branch architecture), Transformer layer fusion, Vocoders (HiFi-GAN, BigV-GAN, UnivNet)

Author countries

South Korea, USA

← Previous