ThinkOmni is a reasoning-driven omni-modal large language model for unified audio forgery reasoning, spoofing detection, and temporal manipulation localization. It combines semantic, acoustic, and spectral-visual evidence to make forensic predictions more explicit and transferable across datasets.
- ThinkOmni jointly produces an inspectable forensic rationale, a three-class authenticity prediction, and one or more manipulated time intervals.
- FACoT contains 100K samples with structured reasoning annotations grounded in semantic inconsistencies, acoustic artifacts, and temporal manipulation patterns.
- FMIL progressively aligns semantic, acoustic, and spectral-visual representations while reducing interference between modalities.
- FCML balances reasoning, detection, and localization objectives with role-aware token weighting and adaptive boundary supervision.
- ThinkOmni reaches 93.70% / 93.72% average intra-dataset ACC / F1 and 80.74% / 85.15% average cross-dataset ACC / F1. For localization, it achieves 88.05% intra-dataset mAP and 74.67% cross-dataset mAP.
The project page is available now. Paper, code, checkpoints, FACoT annotations, and inference examples are being prepared for public release.
| Resource | Status |
|---|---|
| Project page | Available |
| Paper and supplementary material | Coming soon |
| Training and inference code | Coming soon |
| Model checkpoints | Coming soon |
| FACoT annotations | Coming soon |
@misc{xu2026thinkomni,
title = {ThinkOmni: A Reasoning-Driven Omni-Modal LLM Framework for Audio Forgery Detection and Localization},
author = {Yuxiong Xu and Kaiqing Lin and Bin Li and Haodong Li and Sheng Li},
year = {2026}
}ThinkOmni is built on Qwen2.5-Omni, Wav2Vec2 XLS-R, and ms-swift. We thank the authors of these projects and the public audio-forensics benchmarks used in FACoT.