Loading datasets...
Release Date: 11/4/2026
Format: FLAC, TSV
Size: 61.52 GB
Synthetic Hausa (ha, hau) speech for automatic speech recognition, created by CLEAR Global. It contains 622,084 utterances totalling about 993 hours at 24 kHz mono, published as a single train split. The sentences were written by GPT-4o, and the audio was synthesised by a YourTTS model fine-tuned on Hausa Bible recordings (open.bible). Each clip ships with a normalised transcript, the original LLM sentence and its duration. Audio is FLAC (lossless re-encoding of the original 24 kHz WAV).
Licensing
Creative Commons Attribution 4.0 International (CC-BY-4.0)
https://spdx.org/licenses/CC-BY-4.0.htmlThe synthetic text was generated with GPT-4o. Use of the content must comply with OpenAI's Terms of Use and Usage policies, which include not using the outputs for illegal purposes or to generate harmful, violent or infringing content.
None
Ethical Review
The TTS model was fine-tuned on the open.bible Hausa recordings from BibleTTS-Hausa. The narrators are not identified in this release, and no personal data is included. The text is generated and does not contain personal information.
Intended Use
Research and development of automatic speech recognition for Hausa, including training with real data.
This is the Hausa synthetic speech from DeRenzi et al. (2025). In that study, a Wav2Vec-BERT-2.0 model fine-tuned on 250 hours of real plus 250 hours of synthetic Hausa speech matched a 500-hour real-data-only baseline, and 579 hours of real plus 450 to 993 hours of synthetic data gave the best performance.
The sentences were generated with GPT-4o. The audio was synthesised by a YourTTS model fine-tuned by CLEAR Global on Hausa Bible recordings from the BibleTTS dataset.
The linguist-evaluated text sample from the same project (1,600 sentences per language, with readability and adequacy ratings) is published as CLEAR-Global/Synthetic-Text, also on Mozilla Data Collective.
clips/train/ — 622,084 FLAC files, 24 kHz mono, 16-bit, named by the MD5 of the original file name.metadata.tsv — one row per clip with the columns path, split, text, original_text, duration.text is lowercased with punctuation removed and is the transcript to use for ASR training. original_text keeps the LLM's capitalisation and punctuation.
A second synthetic Hausa corpus generated with XTTS-v2 is available as Synthetic Hausa ASR Dataset (XTTS).
DeRenzi, B., Dixon, A., Farhi, M.A., Resch, C. (2025) Synthetic Voice Data for Automatic Speech Recognition in African Languages. arXiv:2507.17578.
@misc{derenzi2025synthetic,
title = {Synthetic Voice Data for Automatic Speech Recognition in African Languages},
author = {DeRenzi, Brian and Dixon, Anna and Farhi, Mohamed Aymane and Resch, Christian},
year = {2025},
eprint = {2507.17578},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2507.17578}
}
This dataset is synthetic: it was not recorded from human speakers. The text was generated by a large language model and the audio by a text-to-speech model, so it may contain model-induced biases. The work was supported by the Gates Foundation; the conclusions and opinions expressed are those of the authors alone and shall not be attributed to the Foundation.
Please check README.md for more information.