Loading datasets...
Release Date: 11/4/2026
Format: FLAC, TSV
Size: 5.94 GB
A speech corpus of 7,594 short Yoruba audio clips with matching transcripts, built for training automatic speech recognition and text-to-speech systems. Clips are 1 to 30 seconds long (median 11.4 s, mean 11.9 s), recorded in studio conditions at 48 kHz, 24-bit, mono FLAC, and each one pairs with its exact transcript. The source is the Yoruba portion of BibleTTS, published by Masakhane from the Open.Bible recordings and the Open Contemporary Bible translation released by Biblica; this release repackages it unchanged. Two narrators read different sets of books. Clips are published in three splits: train (7,491 clips, 24.77 h), dev (63 clips, 0.23 h) and test (40 clips, 0.13 h), with dev and test held out at book level.
Licensing
Creative Commons Attribution Share Alike 4.0 International (CC-BY-SA-4.0)
https://spdx.org/licenses/CC-BY-SA-4.0.htmlNone
None
Ethical Review
The audio derives from the publicly released Open.Bible recordings of Biblica, published openly through the BibleTTS corpus. The two narrators are not identified anywhere in the release and no personal data is included. The recordings and text are openly licensed by the publisher for redistribution and reuse under CC BY-SA 4.0.
Intended Use
Training and evaluation of text-to-speech and automatic speech recognition systems for Yoruba.
BibleTTS is a Masakhane project: the BibleTTS project page documents the corpus and hosts the TTS models and audio samples. The recordings and text come from the Open.Bible project: chapter-level studio recordings plus the Open Contemporary Bible translation, both released by Biblica. BibleTTS cut the chapter-long recordings into short clips and force-aligned them to the text, so every clip is a brief utterance with an exact transcript — usable directly for ASR and TTS training, not as continuous audiobook material.
Yoruba was one of the two languages (with Ewe) aligned from scratch in BibleTTS: an acoustic model was trained per language and the data force-aligned with the Montreal Forced Aligner, using graphemes in place of a grapheme-to-phoneme dictionary, followed by transcript/audio length-ratio outlier filtering tuned per language. Native-speaker annotators judged 76% of a 50-clip sample to be exact matches between audio and transcript.
clips// — 7,594 FLAC files, 48 kHz, 24-bit, mono, named __Verse_.flac.metadata.tsv — one row per clip with the columns path, split, book, text, duration.| split | clips | hours |
|---|---|---|
| train | 7,491 | 24.77 |
| dev | 63 | 0.23 |
| test | 40 | 0.13 |
| total | 7,594 | 25.13 |
dev is the book of Ezra and test is Colossians; both are held out at book level, so no clip of those books appears in train.
The audio is unchanged: same FLAC files, same 48 kHz 24-bit mono format, not resampled or re-aligned. Only the directory layout (book folders folded into the filenames) and the added metadata.tsv differ from the source corpus.
The corpus is not a continuous reading of the Bible. It holds the clips that came out cleanly and survived alignment-quality filtering: 7,594 clips from 1,078 chapters across all 66 books, a median of 8 clips per chapter — both testaments, poetry and prose.
The corpus has two narrators, split by book rather than interleaved. One voice reads Genesis–Deuteronomy, 1–2 Chronicles, Ezra, Nehemiah, Esther, Isaiah, Jeremiah, Lamentations, Ezekiel, Daniel and the entire New Testament; the other reads Joshua–2 Kings (including Ruth), Job, Psalms, Proverbs, Ecclesiastes, Song of Songs and the twelve minor prophets. No clip mixes the two. The source corpus carries no speaker labels, so none were added.
The Interspeech 2022 paper reports 10,228 aligned Yoruba samples (33.3 h) for this corpus. The release available today contains 7,594 clips (25.1 h), and the mean clip length is the same in both (11.7 s vs 11.9 s), so this is a further removal of clips — alignment-quality filtering applied after the paper — rather than a different segmentation. The corpus as distributed is the authoritative version and is what this release reproduces.
Meyer, J., Adelani, D., Casanova, E., Öktem, A., Whitenack, D., Weber, J., KABONGO KABENAMUALU, S., Salesky, E., Orife, I., Leong, C., Ogayo, P., Chinenye Emezue, C., Mukiibi, J., Osei, S., AGBOLO, A., Akinode, V., Opoku, B., Samuel, O., Alabi, J., Muhammad, S.H. (2022) BibleTTS: a large, high-fidelity, multilingual, and uniquely African speech corpus. Proc. Interspeech 2022, 2383-2387, doi: 10.21437/Interspeech.2022-10850
@inproceedings{meyer22c_interspeech,
title = {{BibleTTS: a large, high-fidelity, multilingual, and uniquely African speech corpus}},
author = {Josh Meyer and David Adelani and Edresson Casanova and Alp Öktem and Daniel Whitenack and Julian Weber and Salomon {KABONGO KABENAMUALU} and Elizabeth Salesky and Iroro Orife and Colin Leong and Perez Ogayo and Chris {Chinenye Emezue} and Jonathan Mukiibi and Salomey Osei and Apelete AGBOLO and Victor Akinode and Bernard Opoku and Olanrewaju Samuel and Jesujoba Alabi and Shamsuddeen Hassan Muhammad},
year = {2022},
booktitle = {{Interspeech 2022}},
pages = {2383--2387},
doi = {10.21437/Interspeech.2022-10850},
issn = {2958-1796},
}
The recordings and text are derived from the Open.Bible project by Biblica and are redistributed here as a derivative work of the Masakhane BibleTTS corpus, with the agreement of the corpus authors. CLEAR Global is a partner in the BibleTTS initiative; it did not create, align or modify this audio, only repackaged it. Bible-domain speech by two narrators: not representative of everyday spoken Yoruba. Provided as-is for research and development.
Please check README.md for more information.