MooreSpeechCorporaV2 is the second version of MooreSpeechCorpora, which collects publicly available Mooré speech, aligned with its text, for speech recognition and text-to-speech research. This version holds tales, riddles, proverbs and the Mooré Old Testament. Requests are reviewed by hand, usually within a few days. The recordings come from third-party publications: Mooré oral literature (mooreburkina.com and its ipsapps pages, published by Urs Niggli) and the Mooré Bible read on bible.com (Moore Bible © Bible Society of Burkina Faso, 2014). They are shared for non-commercial research only and remain the property of their owners. License: CC BY-NC 4.0 for our annotations only; the audio and texts stay the property of their owners. Please read the dataset card, its terms and licenses before requesting access.

MooreSpeechCorporaV2

Publicly available Mooré speech, re-aligned and checked, for text-to-speech and speech recognition.

En bref. Deuxième version de MooreSpeechCorpora : contes, devinettes et proverbes réalignés, sans la musique en début et fin d'extrait, vérifiés par deux modèles de reconnaissance vocale (6 306 extraits, 5,4 h). Depuis la v2.2, le sous-ensemble bible ajoute l'Ancien Testament (53 765 extraits, 62,5 h). Accès sur demande, pour la recherche non commerciale.

MooreSpeechCorporaV2 is the second version of MooreSpeechCorpora, Mooré (ISO 639-3 mos) speech paired with its text. Every clip was aligned again, music at the edges was cut, every transcript was checked by two ASR models and the speakers were labelled again. The first six columns are V1's, so V1 code works unchanged. It was built with FasoSpeechCorpora.

Access is gated and requests are reviewed by hand. Collecting and cleaning Mooré speech takes a lot of time and resources, and we do it mostly out of passion. Please read this card, its terms and licenses before requesting access, and make sure you understand them. Respecting them keeps this work useful to the community and helps move our local languages forward. Anonymous or vague requests will not be approved.

Overview

Subset	Content	Clips	Hours	Speakers
`default`	`contes` + `devinettes` + `proverbes`	6,306	5.43	1, 2, 3
`contes`	Tales	4,501	3.79	1 (F), 3 (M)
`devinettes`	Riddles	457	0.31	1 (F)
`proverbes`	Proverbs, Mooré readings	1,348	1.33	1 (F), 2 (M)
`bible`	Old Testament, one clip per sentence	53,765	62.47	4, 5, 6 (M), 7 (F)
`asr`	All of the above for ASR: edge jingles cut, music under the voice kept	60,734	68.35	1 to 7
`asr_tagged`	`asr` clips untrimmed, `<MUSIC>` on edges with at least 0.5 s of music	60,734	72.77	1 to 7
`asr_negatives`	Music-only clips with empty text (52 riddle jingles, 6 Bible clips)	74	0.09	none
`fra`	French readings of the proverbs	796	1.13	1 (F), 2 (M)

Audio is 16 kHz mono. Median CER is 0.050 on default (better of two ASR models per clip) and 0.059 on bible (facebook/mms-1b-all alone, see Limits). default leaves out the Bible, which would be 92% of its hours, from one domain and mostly one reader. Earlier versions: revision="v2.1" (before the Bible) and revision="v2.0" (first release, 5,326 clips).

Quick start

from datasets import load_dataset

ds = load_dataset("anyantudre/MooreSpeechCorporaV2", split="train")             # default (TTS)
asr = load_dataset("anyantudre/MooreSpeechCorporaV2", "asr", split="train")     # one subset
bible = load_dataset("anyantudre/MooreSpeechCorporaV2", "bible", split="train") # the Bible (about 14 GB)

Columns

Column	Type	Description
`audio`	Audio (16 kHz)	The clip
`text`	string	Transcript (Mooré; French in `fra`; empty in `asr_negatives`)
`duration`	float	Length in seconds
`voice`	string	`male` or `female`
`speaker_id`	int	1 to 3: tales, riddles, proverbs; 4 to 7: Bible
`denoised_audio`	Audio (16 kHz)	Voice track without music (Demucs `htdemucs`). TTS subsets only
`source`	string	Hub dataset and volume, ipsapps book URL, or `local`
`chapter`	string	Tale, page or chapter (Bible: e.g. `GEN_1_MPBU`)
`alignment_score`	float	MMS_FA score; empty for clips re-cut from V1
`cer`	float	CER of the transcript against the better of two ASR outputs
`category`	string	`conte`, `devinette`, `proverbe` or `bible`
`book`	string	Source book, e.g. `Yel-bũna vol. 2`
`id`	string	Stable clip id
`language`	string	`mos` or `fra`
`segment_type`	string	`speech`, `speech_with_edge_music` or `music_only`
`is_nonspeech`	bool	True for music-only clips
`nonspeech_spans`	list	What trimming cut at each edge: `start`, `end` (s, untrimmed clip), `label` (`music` or `silence`)
`text_tagged`, `token_lang`, `cmi`, `n_switch_points`, `cs_type`, `matrix_lang`		Code-switching marks (French in `<fr>...</fr>`, language per token, mixing index). Only 7 clips have any

Sources and credits

Tales, riddles and proverbs were published by Urs Niggli on mooreburkina.com, its pages on media.ipsapps.org/mos/ora/, and the "Sagesse Africaine" and "Burkina Langues" apps. Clips re-cut from V1 come from its contes and devinettes subsets.

Book	Authors and readers	Clips	Hours	Speaker
V1 `contes` vol3 to vol5, re-cut; vol3 is probably Contes mooré vol. 1	vol3 probably by Jacques Siballo, read by Ruth Ouedraogo; vol4 and vol5 not identified	4,061	2.98	1 (F)
Contes mooré vol. 2, 18 tales, aligned from the full recordings	Jacques Siballo	440	0.80	3 (M)
V1 `devinettes`, re-cut (probably Devinettes en mooré)	Not credited	457	0.31	1 (F)
Proverbes en mooré vol. 1 to 8 (`p1` to `p8`)	French translation by Timothée Kaboré	610	0.54	2 (M)
Yel-bũna vol. 1 to 3 (`prv-v09` to `prv-v11`)	Attributed to Alain Joseph Sissao (audio file names)	738	0.78	1 (F)

bible is the Old Testament of Wẽnnaam Sebre 2014 (MPBU) on bible.com, from its 929 chapter recordings (74 h). Text © Bible Society of Burkina Faso, 2014; the audio producer and readers are not named. One reader per book:

Speaker	Books	Clips	Hours
4 (M)	Genesis to Deuteronomy, Psalms, Proverbs, Ecclesiastes, Song of Songs, Isaiah, Jeremiah, Lamentations	30,891	30.31
5 (M)	Joshua, Judges, 1 Samuel to 2 Chronicles, Job	13,595	20.89
6 (M)	Ezra, Nehemiah, Ezekiel, Daniel, the twelve minor prophets	8,845	10.39
7 (F)	Ruth, Esther	434	0.88

We thank Urs Niggli, the authors and the readers for making Mooré oral literature available online, and the Bible Society of Burkina Faso for the Mooré Bible. Tools: Meta's MMS with uroman, Demucs, facebook/mms-1b-all, anyantudre/moore-asr-eval-w2vbert and GlotLID.

How it was built

Recipes: recipes/moore-v2/. For each source:

Prepare: 16 kHz mono, clean text. V1 tales vol5, badly cut, were first joined back into one recording per chapter.
Align: MMS_FA with uroman.
Trim: cut music and silence at the edges (Demucs voice track).
Verify: transcribe with both ASR models; texts are lowercased, NFC-normalized, untagged, with apostrophes and hyphens as spaces.
Filter: language, CER at most 0.25, 0.8 to 20 s, music under the voice below -30 dB (TTS subsets), no digits, at least 5 characters per second.

Speakers were grouped with ECAPA voice embeddings, by chapter; the Bible clip by clip, then checked against the books (22 clips sit outside their book's reader). The first clip of each Bible chapter, a section heading that sometimes keeps the end of the spoken chapter announcement, was dropped; 97% of the other aligned clips passed the filter.

To evaluate ASR, leave out rows with is_nonspeech true, report how often a model writes text on them, and strip <...> tags first.

What changed since V1

	V1	V2 `default`
Clips	5,020 unique (5,535 with duplicates)	6,306
Hours	7.70	5.43
Speakers	2 labels (two men shared one)	3
Music at clip edges	in 86% of riddle clips	cut, logged in `nonspeech_spans`
Transcript check	none	two ASR models, clips over 0.25 CER dropped

Fewer hours on purpose: French readings moved to fra, 3 to 9 s riddle jingles were cut, and clips with digits, music under the voice or long pauses were dropped. In a blind A/B test on 50 pairs, the listener preferred V2 27 times and V1 never (21 equal, 2 bad in both).

Limits

Speakers: in default, one woman reads 83% of the clips. Fine for single-speaker TTS, narrow for ASR. In bible, one man reads 57%.
Bible: scripture only; at 63 h it outweighs everything else in asr. anyantudre/moore-asr-eval-w2vbert was probably trained on the Mooré Bible, so its bible scores (median CER 0) are not an independent check, and cer mostly reflects that model there. facebook/mms-1b-all gives a median CER of 0.059.
Content: read tales, proverbs, riddles and scripture; no conversation and almost no code-switching.
Benchmark overlap: the bible, asr and asr_tagged subsets contain the bible.com MPBU translation (Old Testament) of 60 of the 84 chapters held out by the Mooré speech benchmark (moore-tts-bench, to become goaicorp/MooreSpeechBench): 16 of 21 dev, 18 of 23 test and 26 of 40 test_unseen chapters, in other wording (3,051 clips, 3.49 h in bible). The dataset has no jw.org audio or text, so none of the benchmark's jw.org test sentences. The benchmark's MPBU (bsoc) dev and test sets are about 98% contained here, so they are not valid tests for models trained on it. A model trained on bible, asr or asr_tagged should say on its card that it has heard the MPBU translation of most moore-tts-bench held-out Old Testament chapters.
Transcripts: some errors remain under the 0.25 CER cut-off; riddles have the highest median CER (0.097).
Pauses: a few riddles keep inner pauses that passed the filter.

Access

Fill in the form on this page. We approve people and institutions that say who they are, describe a research or academic use and accept the non-commercial terms. Questions: Wendyellé A. Alban NYANTUDRE (nyantudre@gmail.com).

License

Audio and source texts: property of their authors, readers and publisher (see Sources and credits). No license is granted; they are shared for non-commercial research and their owners can ask for removal.
Bible (bible, and its clips in asr, asr_tagged and asr_negatives): text © Bible Society of Burkina Faso, 2014; the audio holder is not named. bible.com's terms allow personal use only, and no permission has been granted for this dataset. The Bible clips will be removed if the rights holders ask.
Our work (segmentation, alignment, corrected transcripts, metadata): CC BY-NC 4.0.

For other uses, contact us. By accessing this dataset you agree to these terms.

Citation

@misc{nyantudre2026moorespeechcorporav2,
  title        = {MooreSpeechCorporaV2: a re-aligned and verified Mooré speech corpus for TTS and ASR},
  author       = {Nyantudre, Wendyellé Abubakrh Alban and Bandaogo, Aristide},
  year         = {2026},
  howpublished = {Hugging Face Datasets (gated access)},
  url          = {https://huggingface.co/datasets/anyantudre/MooreSpeechCorporaV2}
}

Contacts

Wendyellé A. Alban NYANTUDRE, curator: nyantudre@gmail.com, LinkedIn
Aristide BANDAOGO: aristide@goaicorporation.org

Downloads last month: 309

Total file size:

32.1 GB