MMS-lab data

license

not-clearly-stated-on-card(Not distributed

access

closed(The paper releases models, not the aligned data.)

dataset_card

no(No card

Meta released the models trained on this data but not the aligned recordings, which come from third-party Bible audio publishers.

• https://arxiv.org/abs/2305.13516 recorded 2026-09-24

  "a new dataset based on readings of publicly available religious texts".

• https://arxiv.org/pdf/2305.13516 recorded 2026-09-24

  "The MMS models are available at https://github.com/pytorch/fairseq/tree/master/examples/mms"; data drawn from "Faith Comes By Hearing, goto.bible and bible.com"; no data download is offered.

The aligned data was never distributed, so there is no download or usage figure to read.

• https://arxiv.org/abs/2305.13516 recorded 2026-09-24

  The MMS paper releases the models trained on the data, not the aligned recordings, so there is no download or user figure to read

MMS-lab spans more languages than any other labeled speech corpus, and Meta trained its MMS recognition and synthesis models for more than a thousand languages on it. It is a step below Mozilla Common Voice because it is one narrow domain, Bible reading, mostly by a single speaker per recording, and the MMS paper found models trained on Common Voice did better on FLEURS.

• https://arxiv.org/abs/2305.13516 recorded 2026-09-24

  "a single multilingual automatic speech recognition model for 1,107 languages, speech synthesis models for the same number of languages".

• https://arxiv.org/pdf/2305.13516 recorded 2026-09-24

  "speech audio paired with corresponding text in 1,107 languages (MMS-lab; 44.7K hours)"; "it is both from a particular narrow domain and most recordings are from a single speaker".