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 and proverbs. Requests are reviewed by hand, usually within a few days. The recordings come from third-party publications of Mooré oral literature (mooreburkina.com and its ipsapps pages, published by Urs Niggli). 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.

MooreSpeechCorporaV2

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

En bref. Deuxième version de MooreSpeechCorpora, qui rassemble la parole en mooré disponible publiquement, alignée avec son texte. Cette version contient des contes, devinettes et proverbes réalignés, sans la musique en début et fin d'extrait, et vérifiés par deux modèles de reconnaissance vocale : 6 306 extraits, 5,4 heures. Des sous-ensembles pour la reconnaissance vocale et les lectures en français sont fournis. L'accès est sur demande, pour la recherche non commerciale.

MooreSpeechCorporaV2 is the second version of MooreSpeechCorpora, which aims to collect all publicly available Mooré (ISO 639-3 mos) speech with its text. This version holds read tales, riddles and proverbs. Every clip was aligned again, music at the start and end was cut, every transcript was checked by two Mooré ASR models, and the speakers were labelled again. The first six columns are V1's, so code written for V1 works unchanged.

The same work extends to other languages of Burkina Faso through FasoSpeechCorpora, the toolkit that built this dataset.

Access is gated and requests are reviewed by hand. Anonymous or vague requests will not be approved.

Overview

default: 6,306 clips, 5.43 hours, 16 kHz mono, music trimmed, for TTS and ASR
3 speakers: one woman and two men
Median CER 0.050 (lower of the two ASR models' scores for each clip)

Subset	Content	Clips	Hours	Speakers
`default`	`contes` + `devinettes` + `proverbes`	6,306	5.43	1, 2, 3
`contes`	Tales	4,501	3.79	1 (female), 3 (male)
`devinettes`	Riddles	457	0.31	1 (female)
`proverbes`	Proverbs, Mooré readings	1,348	1.33	1 (female), 2 (male)
`asr`	Same sources for speech recognition: edge jingles trimmed, music bed under the voice kept	6,437	5.64	1, 2, 3
`asr_tagged`	The `asr` clips untrimmed, with `<MUSIC>` at edges holding at least 0.5 s of music	6,437	7.25	1, 2, 3
`asr_negatives`	Music-only clips with empty text (52 riddle jingles), to teach a model to output nothing	68	0.09	none
`fra`	French readings of the proverbs, by the same two readers	796	1.13	1 (female), 2 (male)

The version first published, with 5,326 clips in one train split, stays available with revision="v2.0".

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

Columns

Column	Type	Description
`audio`	Audio (16 kHz)	The clip
`text`	string	Transcript (Mooré, or French in `fra`; empty in `asr_negatives`)
`duration`	float	Length in seconds
`voice`	string	Reader's voice: `male` or `female`
`speaker_id`	int	1, 2 or 3
`denoised_audio`	Audio (16 kHz)	The clip with music removed (voice track separated by Demucs `htdemucs`). Only in `contes`, `devinettes` and `proverbes`
`source`	string	Where the clip comes from: Hub dataset and volume, ipsapps book URL, or `local`
`chapter`	string	Tale, page or chapter inside the source
`alignment_score`	float	Forced-alignment score (MMS_FA); empty for clips re-cut from V1
`cer`	float	Character error rate of the transcript against the better of two ASR outputs
`category`	string	`conte` (tale), `devinette` (riddle) or `proverbe` (proverb)
`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 the trim step cut at each edge: `start`, `end` (seconds in the untrimmed clip) and `label` (`music` or `silence`)
`text_tagged`, `token_lang`, `cmi`, `n_switch_points`, `cs_type`, `matrix_lang`		Code-switching annotation (French words in `<fr>...</fr>`, language per token, mixing index). Only 7 clips have any; the columns are there for future sources

Sources and credits

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

Book	Authors and readers	Clips	Hours	Speaker
Tales from MooreSpeechCorpora `contes` vol3 to vol5, re-cut. Books not identified; 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 (female)
Contes mooré vol. 2, 18 tales, aligned again from the full recordings	Written by Jacques Siballo	440	0.80	3 (male)
Riddles from MooreSpeechCorpora `devinettes`, re-cut (probably Devinettes en mooré)	Not credited on the site	457	0.31	1 (female)
Proverbes en mooré vol. 1 to 8 (`p1` to `p8`)	French translation by Timothée Kaboré	610	0.54	2 (male)
Yel-bũna vol. 1 to 3 (`prv-v09` to `prv-v11`)	Collection attributed to Alain Joseph Sissao (from the audio file names)	738	0.78	1 (female)

We thank Urs Niggli, the authors and the readers for making Mooré oral literature available online. Alignment uses Meta's MMS models with uroman, trimming uses Demucs, and transcripts were checked with facebook/mms-1b-all and anyantudre/moore-asr-eval-w2vbert. French words were detected with GlotLID.

How it was built

The recipes are in recipes/moore-v2/ in FasoSpeechCorpora. Each source went through these steps:

Fetch and prepare: download, convert to 16 kHz mono, clean the text. Tales vol5, badly cut in V1, were first joined back into one recording per chapter, with repeated lines removed.
Align: forced alignment with MMS_FA and uroman.
Trim: cut music and silence at the clip edges using the Demucs voice track. asr keeps the music bed under the voice; asr_tagged keeps the edges and marks them.
Verify: transcribe with the two ASR models above.
Filter: language, CER of 0.25 or less, duration between 0.8 and 20 s, music under the voice below -30 dB (TTS subsets), no digits, at least 5 characters per second.

Before scoring, both texts are lowercased, NFC-normalized and stripped of tags, and apostrophes and hyphens become spaces. The speakers step labelled the readers with ECAPA voice embeddings, grouped by chapter.

Evaluating ASR: leave out rows with is_nonspeech true when computing WER, and report separately how often a model outputs text on them. Strip <...> tags before scoring (the toolkit's score_text does this).

What changed since V1

	V1	V2
Clips	5,020 unique (5,535 before removing duplicates)	6,306
Hours	7.70	5.43
Speakers	2 labels (the two male readers shared one)	3
Music at clip edges	in 86% of riddle clips	cut; what was cut is in `nonspeech_spans`
Transcript check	none	CER against two ASR models, clips over 0.25 dropped

V2 has fewer hours on purpose:

French readings: moved to the fra subset.
Riddles: leading jingles of 3 to 9 s are cut.
Other drops: clips with digits, music under the voice, or long pauses inside.

In a blind A/B listening test on 50 matched clip pairs from the first V2 release, the listener preferred V2 in 27 pairs and V1 in none; 21 pairs were rated the same and 2 were bad in both versions.

Limits

Speakers: three readers, and one woman reads 83% of the TTS clips. Fine for single-speaker TTS, narrow for ASR.
Content: this version has read tales, proverbs and riddles only, with no conversational speech and almost no code-switching.
Transcripts: some still have errors under the 0.25 CER cut-off; riddles have the highest median CER (0.097).
Pauses: a few riddles keep pauses inside that survived the filter.

Access

Fill in the form on this page. Requests are approved for 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 on them; they are shared for non-commercial research and their owners can ask for removal.
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:

2.88 GB