espnet/mms_ulab_v2

MMS ulab v2 is a a massively multilingual speech dataset that contains 8900 hours of unlabeled speech across 4023 languages. In total, it contains 189 language families. It can be used for language identification, spoken language modelling, or speech representation learning.

MMS ulab v2 is a reproduced and extended version of the MMS ulab dataset originally proposed in Scaling Speech Technology to 1000+ Languages, covering more languages and containing more data. This dataset includes the raw unsegmented audio in a 16kHz single channel format. It can be segmented into utterances with a voice activity detection (VAD) model such as this one. We use 6700 hours of MMS ulab v2 (post-segmentation) to train XEUS, a multilingual speech encoder for 4000+ languages.

For more details about the dataset and its usage, please refer to our paper or project page.

License and Acknowledgement

MMS ulab v2 is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license.

If you use this dataset, we ask that you cite the following papers:

@misc{chen2024robustspeechrepresentationlearning,
      title={Towards Robust Speech Representation Learning for Thousands of Languages}, 
      author={William Chen and Wangyou Zhang and Yifan Peng and Xinjian Li and Jinchuan Tian and Jiatong Shi and Xuankai Chang and Soumi Maiti and Karen Livescu and Shinji Watanabe},
      year={2024},
      eprint={2407.00837},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.00837}, 
}

@article{pratap2024scaling,
  title={Scaling speech technology to 1,000+ languages},
  author={Pratap, Vineel and Tjandra, Andros and Shi, Bowen and Tomasello, Paden and Babu, Arun and Kundu, Sayani and Elkahky, Ali and Ni, Zhaoheng and Vyas, Apoorv and Fazel-Zarandi, Maryam and others},
  journal={Journal of Machine Learning Research},
  volume={25},
  number={97},
  pages={1--52},
  year={2024}
}

And also reference The Global Recordings Network, the original source of the data.