CLEAR-Global/BibleTTS-Chichewa

BibleTTS Chichewa — Aligned Bible Speech

116.1 hours of studio-quality, single-speaker Chichewa Bible narration, segmented and force-aligned at verse level into 45,728 clips (24 kHz, mono, WAV).

Languages

• nya — Chichewa or Chewa

Dataset Structure

Audio is stored in the audio column (HuggingFace Audio feature). Parquet shards live under data/read/, one file per split.

from datasets import load_dataset

ds = load_dataset("CLEAR-Global/BibleTTS-Chichewa", "read", split="train")

field	description
audio	The clip. audio.path is the source file name, <BOOK>_<chapter>_<segment>.wav.
text	Verse transcript, lowercase.
duration	Clip duration in seconds.
score	Alignment confidence for the clip (see below).

Statistics

split	clips	hours	median / p95 duration (s)
train	40,150	100.81	8.2 / 17.8
dev	249	0.76	10.5 / 21.0
test	117	0.38	11.2 / 23.7
other	5,212	14.15	7.2 / 26.6
total	45,728	116.10	8.1 / 18.8

The corpus has a single speaker (one recording voice) and no speaker or demographic metadata.

Splits and the Alignment Score

Each clip carries a score from the forced aligner: the alignment confidence for that clip, expressed as an average per-token log-probability. Values are negative; higher is better. It is length-normalised, so long and short clips are comparable, and it is not correlated with clip duration or transcript length.

Every clip is published in one of four splits. Nothing is discarded:

• train / dev / test — the cleaned set (40,516 clips, 101.95 h). A clip is included only if it passes both filters:
  1. score > -45 — alignment considered reliable; and
  2. no numerals in the transcript — Chichewa verse text containing figures ("…wa zaka '32'…") is kept out of the training splits because the numbers are spoken as words and do not match a plain-text transcript.
• other — the 5,212 clips (14.15 h) excluded by those filters; not intended for TTS training as-is.

Dataset Creation

BibleTTS's Chichewa subset was published as unaligned chapter-level audio from Open.Bible released by Biblica. This release is the CLEAR Global verse-level segmentation of those recordings: each chapter was split into per-verse clips and force-aligned against the Chichewa Bible text, which yielded the per-clip transcripts, durations and alignment scores used here. Clips are 24 kHz mono PCM WAV.

Caveats

• Single speaker, Bible-domain vocabulary: not representative of everyday spoken Chichewa.
• Verse-level segmentation follows the Bible text; some clips are verse fragments, so transcripts are not always complete sentences.
• The other split is deliberately unfiltered and therefore contains misaligned clips and transcripts with numerals.

Citation

@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},
}

@inproceedings{derenzi-etal-2025-synthetic,
  title     = "Synthetic Voice Data for Automatic Speech Recognition in {A}frican Languages",
  author    = "DeRenzi, Brian and Dixon, Anna and Farhi, Mohamed Aymane and Resch, Christian",
  booktitle = "Proceedings of the First Workshop on Advancing {NLP} for Low-Resource Languages",
  month     = sep,
  year      = "2025",
  address   = "Varna, Bulgaria",
  publisher = "INCOMA Ltd., Shoumen, Bulgaria",
  url       = "https://aclanthology.org/2025.lowresnlp-1.16/",
  pages     = "152--186"
}

Contact

CLEAR Global — tech@clearglobal.org

Acknowledgments

Verse-level alignment and packaging by CLEAR Global, continuing the Masakhane BibleTTS collaboration. Original recordings from the Open.Bible project by Biblica.

License

CC BY-SA 4.0 (SPDX: cc-by-sa-4.0)