twangodev/librivox-mirror

LibriVox Mirror

Fast, structured, continuously updated LibriVox audio mirror.

Current snapshot

Metric	Value
Published books	14,329
Published sections	325,926
Audio hours	88,415.7
Audio languages	75
Quarantined books	156
Last updated (UTC)	2026-08-31T04:34:55.855597Z

Audio by language

Language	Hours
English	87,723.9
German	351.0
Spanish	89.6
French	73.6
Portuguese	29.0
Dutch	25.8
Italian	17.5
Polish	15.0
Japanese	12.9
Danish	11.6
Latin	6.3
Russian	5.7
Arabic	5.7
Esperanto	4.4
Chinese	3.9
Romanian	3.5
Greek	3.2
Ukrainian	2.3
Finnish	1.8
Ancient Greek	1.7
Multilingual	1.7
Javanese	1.7
Swedish	1.6
Tagalog	1.6
Old English	1.3
Urdu	1.3
Hungarian	1.3
Galician	1.1
Norwegian	1.1
Yiddish	1.0
Persian/Farsi	0.9
Afrikaans	0.8
Indonesian	0.8
Slovenian	0.7
Hebrew	0.7
Bulgarian	0.7
Tamil	0.6
Czech	0.5
Kapampangan	0.5
Latvian	0.5
Middle English	0.5
Luxembourgish	0.4
Balinese	0.4
Malay	0.4
Minangkabau	0.4
Buginese	0.4
Acehnese	0.4
Hindi	0.3
Sundanese	0.3
Catalan	0.3
Oriya	0.3
Nynorsk	0.3
Volapük	0.3
Occitan	0.2
Walloon	0.2
Korean	0.2
Western Frisian	0.2
Croatian	0.2
Slovak	0.2
Turkish	0.2
Faroese	0.2
Bisaya/Cebuano	0.2
Cantonese Chinese	0.2
Old Javanese	0.1
Old Norse	0.1
Bengali	0.1
Scottish Gaelic	0.1
Lithuanian	0.1
Welsh	0.0
Old Sundanese	0.0
Irish	0.0
Serbian	0.0
Low German	0.0
Belarusian	0.0
Old Tupi	0.0

Dataset structure

The default preview config contains a bounded set of browser-playable original MP3s. Full sections and books indexes remain typed Parquet; canonical audio is streamed from WebDataset shards under data/.

Use language and hash_partition to construct stable downstream subsets or evaluation splits. Source metadata features preserve unmodeled LibriVox and Internet Archive fields for provenance.

License and attribution

The copyrightable mirror-specific compilation, curation, normalized metadata, indexes, and documentation are licensed under CC BY 4.0. Please cite this dataset when using those parts. Original LibriVox audio remains public domain in the United States and is not relicensed by this mirror. Rules may differ by jurisdiction.

Citation

@misc{ding2026librivoxmirror,
  author       = {James Ding},
  title        = {LibriVox Mirror},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/twangodev/librivox-mirror}},
  note         = {Continuously updated dataset}
}

Provenance and integrity

Each sample links to its LibriVox project and exact Internet Archive source file. Upstream checksums and a mirror SHA-256 are preserved, and audio is not transcoded.