I acknowledge that this dataset contains synthetic data generated through the process described in this paper and its corresponding GitHub codebase. It is not generated by humans, and as such, may contain model-induced biases.
The synthetic text data in this dataset was generated using GPT-4o. I agree to use the content in accordance with OpenAI’s Terms of Use and Usage policies. This includes not using the outputs for illegal purposes or to generate harmful, violent, or infringing content. I understand that I am responsible for ensuring my use of OpenAI-generated content complies with all applicable laws, regulations, and platform terms.
If the dataset is updated in the future, we will notify users who have provided their contact information.

Synthetic Hausa ASR Dataset (XTTS)

Synthetic Hausa (ha) speech for automatic speech recognition (ASR), created by CLEAR Global. It contains about 574 hours of 24 kHz speech across 490,475 utterances, generated from LLM-written sentences by a text-to-speech (TTS) model fine-tuned on Hausa Bible recordings.

This is a synthetic dataset. The sentences were written by a large language model and the audio was produced by a TTS model, not recorded from human speakers. Models trained on it should be evaluated on real speech before use.

data/train-*.parquet: 201 Parquet shards, one split (train).
audio: WAV audio, 24 kHz, mono, 16-bit PCM, embedded in the Parquet files.
text: normalised transcript, lowercased with punctuation removed. Use this for ASR training.
original_text: the sentence as generated by the LLM, with capitalisation and punctuation.
duration: clip length in seconds.

from datasets import load_dataset

ds = load_dataset("CLEAR-Global/Hausa-Synthetic-ASR-Dataset-XTTS", split="train")

How it was made

Text. Hausa sentences were written by GPT-4o: a corpus of 674,000 sentences (Table 3 of the paper). The linguist-evaluated sample of that corpus (1,600 sentences with readability and adequacy ratings) is published with the other languages as CLEAR-Global/Synthetic-Text, also on Mozilla Data Collective.
Speech. An XTTS-v2 model fine-tuned on Hausa Bible recordings (BibleTTS-Hausa, from open.bible) synthesised the sentences; this release holds the 490,475 clips produced for about 574 hours of audio. Model checkpoint: CLEAR-Global/XTTS-Hausa.

For the full method, evaluation and ASR results, see the paper: DeRenzi, Dixon, Farhi & Resch (2025), Synthetic Voice Data for Automatic Speech Recognition in African Languages, arXiv:2507.17578. For Hausa it reports that Wav2Vec-BERT-2.0 fine-tuned on 250 h of real plus 250 h of synthetic speech matched a 500 h real-data-only baseline, and that 579 h of real plus 450 h to 993 h of synthetic data gave the best performance.

Limitations

Limited voice variety. The TTS model was fine-tuned on the Hausa Bible narrators in BibleTTS-Hausa, so the speech comes from a small number of synthetic voices. No speaker metadata is included, so speaker-conditioned training or evaluation is not possible with this release.
No natural acoustic variation. Real recordings contain tone, noise and speaking-style variation that synthetic speech lacks.
Model-induced bias. The text reflects the LLM's assumptions and may not capture Hausa-speaking communities' culture or everyday language.
Transcripts are the input sentences, not transcriptions of the audio. The text was generated first and then synthesised, so a clip's transcript shows what the TTS model was asked to say. Reviewers rated the text for readability in the paper's evaluation, but the audio was not checked against the transcripts clip by clip.

Text corpus: CLEAR-Global/Synthetic-Text — on MDC
CLEAR-Global/XTTS-Hausa: the TTS model used to generate the audio.
CLEAR-Global/YourTTS-Hausa: Hausa YourTTS model.
CLEAR-Global/BibleTTS-Hausa: real Hausa Bible speech used to fine-tune the TTS model.
CLEAR-Global/Hausa-Synthetic-ASR-Dataset-YourTTS: the YourTTS variant of this corpus (993 h, 622,084 clips).

Funding

This 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.

Citation

@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}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE. The synthetic text was generated with GPT-4o; use of the content must comply with OpenAI's Terms of Use and Usage policies.