Buckets:

---

leonpscl

/

musan-bucket

MUSAN (OpenSLR SLR17) is a corpus of music, speech, and noise recordings — ~109 hours of 16kHz mono audio assembled for training speech-processing systems such as VAD, speaker ID, and noise augmentation. The default config is the full labelled corpus for music/speech/noise discrimination; each subcorpus is also its own config so you can fetch just one:

• music — ~42h across the FMA, HD-classical, Jamendo, and RFM sources.
• speech — ~60h of read speech (LibriVox, 12 languages) plus US government recordings.
• noise — ~6h of technical and ambient noise (free-sound, sound-bible).

from datasets import load_dataset

musan = load_dataset('corypaik/musan', split='train')  # full labelled set
noise = load_dataset('corypaik/musan', 'noise', split='train')  # one subcorpus

License

CC BY 4.0 overall (per the OpenSLR listing), but licensing is mixed by source: free-sound noise is US public domain, sound-bible noise is assorted Creative Commons, and the music and speech sources carry their own terms. Each subcorpus ships a LICENSE mapping every clip to its terms; those files are preserved under licenses/ in this repo.

Homepage: https://www.openslr.org/17/

Citation

@misc{snyder2015musan,
  title={MUSAN: A Music, Speech, and Noise Corpus},
  author={David Snyder and Guoguo Chen and Daniel Povey},
  year={2015},
  eprint={1510.08484},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
}