BEATs: Audio Pre-Training with Acoustic Tokenizers
Paper โข 2212.09058 โข Published
How to use jaeyeonkim99/BEATs-base-AS2M-iter0 with ESPnet:
unknown model type (must be text-to-speech or automatic-speech-recognition)
BEATs base audio encoder pre-trained on AudioSet-2M with ESPnet, reproducing BEATs (Table 4). Iteration 0 is trained against random-projection targets. Produced by the egs2/audioset/ssl1 recipe.
Load into an ESPnet downstream classification recipe (e.g. egs2/esc50/asr1,
egs2/as20k/cls1) via beats_ckpt_path:
encoder: beats
encoder_conf:
beats_ckpt_path: beats_encoder_iter0.pt
fbank_mean: 15.66439
fbank_std: 6.38312
| AS-20K test mAP | ESC-50 (5-fold acc) | |
|---|---|---|
| iter 0 | 29.72 | 93.35 |
@inproceedings{chen2022beats,
title={BEATs: Audio Pre-Training with Acoustic Tokenizers},
author={Chen, Sanyuan and Wu, Yu and Wang, Chengyi and Liu, Shujie and
Tompkins, Daniel and Chen, Zhuo and Wei, Furu},
booktitle={ICML}, year={2023}
}
@inproceedings{bharadwaj2025openbeats,
title={OpenBEATs: A Fully Open-Source General-Purpose Audio Encoder},
author={Bharadwaj, Shikhar and Cornell, Samuele and Choi, Kwanghee and
Fukayama, Satoru and Shim, Hye-jin and Deshmukh, Soham and Watanabe, Shinji},
booktitle={WASPAA}, year={2025}
}