Official implementation: "AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation"

If you think this repo is useful to you, please give it to us ⭐

August 1, 2025: The AudioSet-R paper was accepted by the ACM MM2025 Dataset Track.

May 31, 2025: We submitted our paper on the AudioSet-R Dataset to the ACM MM2025 Dataset track.

Illustrates the proposed three-stage relabeling framework for AudioSet.

The dataset used in this study is AudioSet dataset: Balanced training set (20550 .wav files), Evaluate set (18885 .wav files).

AudioSet official website: AudioSet, Github

We provide the json file of the audioset-R audio label and the json file of the original AudioSet audio label:

👍AudioSet-R: The json file of balanced training set and evaluate set

Original AudioSet: The json file of balanced training set and evaluate set

The detailed analysis for three-round audio content extraction.

We provide a comparison of the training and fine-tuning performance of AudioSet-R on various supervised and self-supervised audio classification models:

If you want to perform label prediction on other datasets, please follow the steps below:

cd ./pipeline

Please read ./pipeline/README.md

If you found this repo is helpful, please consider citing our papers:

@article{sun2025audiosetr,
  title={AudioSet-R: A Refined AudioSet with Multi-Stage LLM Label Reannotation},
  author={Yulin Sun and Qisheng Xu and Yi Su and Qian Zhu and Yong Dou and Xinwang Liu and Kele Xu},
  journal={arXiv preprint arXiv:2508.15429},
  year={2025}
}