MooreSpeechBench is the evaluation protocol of a research project on Mooré text-to-speech and speech recognition: test-set ids, rules, scoring code and results. The audio and transcripts it points to come from third-party Bible audio and stay in their own gated repositories; they are for non-commercial research only. Our lists, protocol, code and results are CC BY-NC 4.0. Commercial use needs written permission from nyantudrealban@gmail.com. Please read the card, its terms and licenses before requesting access, and make sure you understand them.

MooreSpeechBench

Frozen test sets, rules and scoring code for Mooré text-to-speech and speech recognition.

MooreSpeechBench is our own evaluation protocol for Mooré speech. No public Mooré speech test exists: FLEURS, Common Voice, WAXAL and BibleTTS have no Mooré. So we built one from held-out chapters of read Bible audio and wrote down how the sets were made, what a scored model may train on, and how scores are computed. This repository holds the test-set ids, the protocol, the normaliser and scoring code, and the v1.0 results. The audio and transcripts stay in anyantudre/moore-tts-bench, pinned at commit f0665e2.

Résumé en français. MooreSpeechBench est notre protocole d'évaluation pour la parole en mooré (synthèse et reconnaissance vocales). Il contient les identifiants des jeux de test, figés : test, 541 extraits de voix vues à l'entraînement, et test_unseen, 836 extraits de deux voix nouvelles. Il contient aussi les règles d'entraînement, la normalisation du texte, le code de notation et les résultats de la version 1.0. L'audio et les transcriptions restent dans anyantudre/moore-tts-bench (commit f0665e2), un dépôt à accès contrôlé. Les scores sont des taux d'erreur sur les caractères (CER) mesurés par deux juges de reconnaissance vocale, avec des intervalles de confiance à 95 % par rééchantillonnage des chapitres. Licence CC BY-NC 4.0, évaluation seulement : ne pas entraîner de modèle sur ces jeux. Accès sur demande.

Evaluation only. Do not train or fine-tune on any set listed here, in any source or translation.

Sets

Set	Clips	Hours	What it measures	Role
`test`	541	1.46	Voices seen in training (9 readers), on unseen chapters	Reported
`test_unseen`	836	2.53	Two new voices (V17, V18) on new books	Reported
`dev`	402	1.03	Seen voices	Model selection only
`dev150`	150	0.37	Subset of `dev` with frozen voice prompts	Checkpoint selection only

All four sets are jw.org audio (the jw24 part of moore-tts-bench). The lists are in sets/: ids, chapter keys, readers, durations and frozen prompt ids, with no text. The references come from moore-tts-bench by id. sets/heldout_chapters.csv lists every held-out chapter. sets/near_copies.csv flags the 32 verses that copy, or nearly copy, a training verse (24 of them in test).

Additional ASR sets (not part of the v1.0 headline):

Set	Clips	Hours	Source
`bc_heldout` dev / test / test_unseen	713 / 957 / 1,403	3.50 in all	MooreSpeechCorporaV2 v2.3 `asr` clips of the same held-out chapters, read by the bible.com readers
`goai_test`	366	0.47	Non-Bible read sentences, 12 readers. It needs access to a private corpus.

How to score

The full rules are in PROTOCOL.md.

Judges. Two speech recognisers that never heard a held-out chapter in any source.
- anyantudre/mos-w2v-bert-2.0-bench, revision b834419, is judge 1.
- anyantudre/mos-w2v-bert-2.0-jw, revision fafc2ce, is judge 2. It was trained on jw.org audio only and has heard the same readers, so it is a robustness check, not an independent verdict.
- Both use greedy CTC decoding with no language model.
Normaliser. moore-asr-norm 1.0 (code/mtts/text.py, normalize_cer): lowercase, apostrophes and hyphens become spaces, and only letters, combining marks and spaces are kept. Tildes are never stripped.
TTS. Synthesise each set with its frozen prompts and 3 synthesis seeds. The headline is the paired gap to the real recordings of the same verses, ΔCER = CER(system) − CER(real), with a 95 % chapter bootstrap (2,000 resamples). Speaker similarity (ECAPA) and UTMOSv2 are reported as diagnostics.
ASR. CER first, then WER. test and test_unseen are never pooled.
Code. code/mtts/eval (scoring), code/tools/test_report.py (TTS tables) and code/server/eval_asr_sets.py (ASR sets).

v1.0 results

TTS, judge 1, 3 seeds. The real recordings score 0.54 % CER on test and 0.73 % on test_unseen. Every gap below is beyond 2 standard errors, except where the note says otherwise.

System	`test` CER	ΔCER [95 % CI]	`test_unseen` CER	ΔCER [95 % CI]
Kokoro	0.38 %	−0.16 [−0.24, −0.07]	0.43 %	−0.30 [−0.39, −0.21]
OmniVoice	0.90 %	+0.35 [+0.12, +0.70]	1.07 %	+0.34 [+0.15, +0.56]
Chatterbox	1.30 %	+0.75 [+0.56, +1.00]	1.44 %	+0.71 [+0.54, +0.91]
Spark-TTS	1.55 %	+1.00 [+0.74, +1.30]	1.50 %	+0.77 [+0.61, +0.93]
Pocket TTS teacher	1.61 %	+1.07 [+0.70, +1.55]	1.89 %	+1.16 [+0.85, +1.46]
Pocket TTS student	1.89 %	+1.34 [+0.82, +1.99]	2.67 %	+1.94 [+1.34, +2.53]
F5-TTS	1.95 %	+1.40 [+1.20, +1.64]	2.33 %	+1.60 [+1.46, +1.74]
MMS-TTS (VITS)	13.48 %	+12.93 [+12.05, +13.88]	13.66 %	+12.93 [+12.36, +13.44]

Judge 2 ranks the eight systems in the same order on both sets. Its scores are a little higher (real recordings 0.63 % and 0.90 %). Under judge 2, OmniVoice's gap on test is +0.31, between 1 and 2 standard errors. Kokoro and MMS-TTS cannot clone a voice, so on test_unseen they speak with the nearest training voice. Full tables, with speaker similarity and UTMOSv2: results/v1_test_results.md and results/v1_test_results_judge2.md.

ASR, real recordings, CER.

Model	`test`	`test_unseen`	`bc_heldout` test	`goai_test`
Judge 1	0.54 %	0.73 %	1.18 %	9.36 %
Judge 2	0.63 %	0.90 %	3.09 %	9.97 %
Our all-data w2v-BERT 2.0 (137.6 h)	0.57 %	0.75 %	1.16 %	6.46 %
facebook/mms-1b-all (Mooré adapter)	6.44 %	6.34 %	6.38 %	16.00 %

Details and confidence intervals are in results/asr_all_results.md.

Training-data rules

A scored model must not train on any held-out chapter, in any source or translation. This includes text used by a speech system, such as a decoding language model.
Models trained on MooreSpeechCorpora V1 or V2 Bible data have heard the bible.com version of 60 held-out Old Testament chapters. They may be scored, with a note on their card and the exposure test in PROTOCOL.md.
Choose checkpoints on dev or dev150 only. Run test and test_unseen once per release candidate.

How it was built

The sets come from moore-tts-bench (manifest_v1, split seed 20261024):

Whole chapters are held out per split.
Readers are clusters of speaker embeddings (V01 to V20).
Verses are flagged as near-copies when they match a training verse exactly or share at least 80 % of their 4-grams with the training text.

Limits

Read Bible speech only. test uses the training voices, and test_unseen mixes new voices with new books.
The best TTS systems sit close to the real-recording floor. Compare them with the paired test, not with raw CER.
Rebuilding the sets needs access to the gated source repositories.

Versions

v1.0: the splits of moore-tts-bench at commit f0665e2, a judge retrained without any held-out chapter, and a second judge.
Planned v2: a book-level held-out set, MooreSpeechCorporaV2 as the bible.com source, and jw.org data rebuilt with the same toolkit. v1 and v2 scores will not be comparable.

License

The lists, protocol, code and results are our own work, under CC BY-NC 4.0. Commercial use needs written permission from nyantudrealban@gmail.com. The audio and transcripts belong to their owners and are not part of this repository.

Citation

@misc{nyantudre2026moorespeechbench,
  title        = {MooreSpeechBench: an evaluation protocol for Mooré text-to-speech and speech recognition},
  author       = {Nyantudre, Wendyellé Abubakrh Alban},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/anyantudre/MooreSpeechBench}
}

Downloads last month: 31

Total file size:

611 kB