Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fusion Segment Transformer: Bi-Directional Attention Guided Fusion Network for AI-Generated Music Detection

Yumin Kim, Seonghyeon Go

TL;DR

Addresses the challenge of detecting AI-generated music in full-length tracks by modeling long-term musical structure. Proposes Fusion Segment Transformer, a two-stage framework that fuses short-segment content embeddings with global structural cues via a bi-directional cross-attention fusion layer and a learnable gating mechanism. Key contributions include a flexible, extractor-agnostic fusion architecture, cross-modal interaction between content and structure, and state-of-the-art results on the SONICS and AIME datasets. The approach improves robustness to generation artifacts and long-range musical regularities, with practical implications for copyright enforcement and music forensics.

Abstract

With the rise of generative AI technology, anyone can now easily create and deploy AI-generated music, which has heightened the need for technical solutions to address copyright and ownership issues. While existing works mainly focused on short-audio, the challenge of full-audio detection, which requires modeling long-term structure and context, remains insufficiently explored. To address this, we propose an improved version of the Segment Transformer, termed the Fusion Segment Transformer. As in our previous work, we extract content embeddings from short music segments using diverse feature extractors. Furthermore, we enhance the architecture for full-audio AI-generated music detection by introducing a Gated Fusion Layer that effectively integrates content and structural information, enabling the capture of long-term context. Experiments on the SONICS and AIME datasets show that our approach outperforms the previous model and recent baselines, achieving state-of-the-art results in AI-generated music detection.

Fusion Segment Transformer: Bi-Directional Attention Guided Fusion Network for AI-Generated Music Detection

TL;DR

Addresses the challenge of detecting AI-generated music in full-length tracks by modeling long-term musical structure. Proposes Fusion Segment Transformer, a two-stage framework that fuses short-segment content embeddings with global structural cues via a bi-directional cross-attention fusion layer and a learnable gating mechanism. Key contributions include a flexible, extractor-agnostic fusion architecture, cross-modal interaction between content and structure, and state-of-the-art results on the SONICS and AIME datasets. The approach improves robustness to generation artifacts and long-range musical regularities, with practical implications for copyright enforcement and music forensics.

Abstract

With the rise of generative AI technology, anyone can now easily create and deploy AI-generated music, which has heightened the need for technical solutions to address copyright and ownership issues. While existing works mainly focused on short-audio, the challenge of full-audio detection, which requires modeling long-term structure and context, remains insufficiently explored. To address this, we propose an improved version of the Segment Transformer, termed the Fusion Segment Transformer. As in our previous work, we extract content embeddings from short music segments using diverse feature extractors. Furthermore, we enhance the architecture for full-audio AI-generated music detection by introducing a Gated Fusion Layer that effectively integrates content and structural information, enabling the capture of long-term context. Experiments on the SONICS and AIME datasets show that our approach outperforms the previous model and recent baselines, achieving state-of-the-art results in AI-generated music detection.
Paper Structure (16 sections, 3 equations, 3 figures, 2 tables)

This paper contains 16 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Segment Transformer
  4. Methods
  5. Stage-1: Feature Embedding Extractor for Short Audio Segment Detection
  6. Stage-2: Fusion Segment Transformer for Full Audio Segment Detection
  7. Dual-Stream Processing
  8. Cross-Modal Fusion Layer
  9. Experiments and Results
  10. Datasets
  11. Training Details and Evaluation Settings
  12. Evaluation Results
  13. Ablation Study
  14. Impact of Segmentation Strategy
  15. Effect of the Gated Fusion Layer
  16. ...and 1 more sections

Figures (3)

  • Figure 1: The architecture of Stage-1 for short-audio segment detection. Various feature extractors (a–e) are flexibly selected on the Encoder of AudioCAT framework.
  • Figure 2: Stage-2 fusion-enhanced Segment Transformer architecture with dual-stream processing, cross-modal fusion mechanism, and multi-scale adaptive pooling for full-audio AIGM detection.
  • Figure 3: Visualization of fusion gate dynamics for Real and Fake tracks on the AIME dataset grotschla2025benchmarking.