espnet/yodas_owsmv4

This dataset comprises 166,000 hours of multilingual speech spanning 75 languages, segmented into 30-second long-form audio clips. The data originates from the YODAS2 dataset, which is based on large-scale web-crawled content. Due to the nature of web-sourced data, the original YODAS2 dataset may include inaccurate language labels and misaligned audio-text pairs. Our preliminary experiments indicate that such noise can negatively impact downstream ASR model performance. To address this, we developed a scalable data-cleaning pipeline using publicly available toolkits, resulting in a curated subset of the original dataset. This cleaned dataset forms a core part of the training data for our OWSM v4 models, which, when combined with existing OWSM data, significantly outperform previous versions on multilingual ASR benchmarks.

YODAS Data Cleaning

Please refer to our paper for more details: https://arxiv.org/abs/2506.00338

OWSM v4 series

Encoder-decoder OWSM

Name	Size	Hugging Face Repo
OWSM v4 base	102M	https://huggingface.co/espnet/owsm_v4_base_102M
OWSM v4 small	370M	https://huggingface.co/espnet/owsm_v4_small_370M
OWSM v4 medium	1.02B	https://huggingface.co/espnet/owsm_v4_medium_1B

CTC-based OWSM

Name	Size	Hugging Face Repo
OWSM-CTC v4 medium	1.01B	https://huggingface.co/espnet/owsm_ctc_v4_1B

Citation

@inproceedings{owsm-v4,
  title={{OWSM} v4: Improving Open Whisper-Style Speech Models via Data Scaling and Cleaning},
  author={Yifan Peng and Shakeel Muhammad and Yui Sudo and William Chen and Jinchuan Tian and Chyi-Jiunn Lin and Shinji Watanabe},
  booktitle={Proceedings of the Annual Conference of the International Speech Communication Association (INTERSPEECH) (accepted)},
  year={2025},
}