Technical Report of Nomi Team in the Environmental Sound Deepfake Detection Challenge 2026

TL;DR

This paper describes the Nomi Team's approach to the 2026 Environmental Sound Deepfake Detection Challenge, using an audio-text cross-attention model to improve detection of synthetic environmental sounds.

Contribution

It introduces a novel audio-text cross-attention model tailored for environmental sound deepfake detection, addressing unseen generators and low-resource scenarios.

Findings

Achieved competitive EER improvements over baseline models

Demonstrated effectiveness of combined text-audio models

Addressed challenges of unseen sound generators

Abstract

This paper presents our work for the ICASSP 2026 Environmental Sound Deepfake Detection (ESDD) Challenge. The challenge is based on the large-scale EnvSDD dataset that consists of various synthetic environmental sounds. We focus on addressing the complexities of unseen generators and low-resource black-box scenarios by proposing an audio-text cross-attention model. Experiments with individual and combined text-audio models demonstrate competitive EER improvements over the challenge baseline (BEATs+AASIST model).