anyantudre/mos-w2v-bert-2.0-bench
Automatic Speech Recognition • 0.6B • Updated • 36
Datasets:
Request access to moore-tts-bench
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
moore-tts-bench is the frozen benchmark of a research project on Mooré text-to-speech. The recordings come from third-party Bible audio (listed on the card) and are shared for non-commercial research only; our annotations are CC BY-NC 4.0. Please read the card, its terms and licenses before requesting access, and make sure you understand them.
Log in or Sign Up to review the conditions and access this dataset content.
Mooré TTS benchmark (gated, research use)
Preprocessed from anyantudre/moore-speech-bible with a frozen manifest (manifest/manifest_v1.parquet,
seed 20261024). Mono, edge silences trimmed to 0.15 s, loudness normalised to -20 LUFS, FLAC 16 bit.
jw24: jw.org translation, 24 kHz. Splits: train / dev / test (held-out chapters) / test_unseen (voices V17, V18).bsoc: Bible-society translation (bible.com 16 kHz + jw.org GEN-NUM resampled to 24 kHz), ablation only.- Held-out chapters are removed from every source. Nested data budgets: columns
b1h_s0,b5h_s0,b20h_s0,b_all(+ seeds 1-2 at 1 h and 5 h). - Voice ids come from speaker-embedding clustering of chapters;
qc_cer_mmsis a quality-control score only.
| config | split | clips | hours |
|---|---|---|---|
| bsoc | dev | 626 | 0.69 |
| bsoc | test | 822 | 0.99 |
| bsoc | train | 48171 | 65.45 |
| jw24 | dev | 402 | 1.03 |
| jw24 | test | 541 | 1.46 |
| jw24 | test_unseen | 836 | 2.53 |
| jw24 | train | 20491 | 54.65 |
The source audio (jw.org, bible.com) belongs to its owners and its terms restrict redistribution. Access is granted by manual review, for non-commercial research only, and you may not redistribute the audio or the transcripts.
- Downloads last month
- 109