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 Claude Sonnet 3.5. I agree to use the content only for non-commercial purposes, and in accordance with Anthropic’s Terms of Service and Usage Policy. This includes not using the outputs to develop products or services that compete with Anthropic’s offerings.
If the dataset is updated in the future, we will notify users who have provided their contact information.

Synthetic Chichewa ASR Dataset

Synthetic Chichewa (ny) speech for automatic speech recognition (ASR), created by CLEAR Global. It contains about 550 hours of 24 kHz speech across 374,612 utterances, generated from LLM-written sentences by a text-to-speech (TTS) model fine-tuned on Chichewa 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. The voice is a single male voice. Models trained on it should be evaluated on real speech before use.

data/train-*.parquet: 191 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/Chichewa-Synthetic-ASR-Dataset", split="train")

How it was made

Text. Short Chichewa sentences and questions were generated with Claude 3.5 Sonnet, which the authors found to give the most natural Chichewa of the models they tested. The large corpus contained many duplicates (only 37% of 700,000 generated sentences were unique), so the deduplicated corpus was used.
Speech. A YourTTS model fine-tuned on Chichewa Bible recordings (open.bible) synthesised each sentence. Model checkpoint: CLEAR-Global/YourTTS-Chichewa.

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. The Chichewa portion of the paper reports about a 6.5% relative WER improvement when training on a 1:2 mix of real and synthetic data.

Limitations

Single male voice. The TTS model was fine-tuned on male speakers only, so the synthetic speech is exclusively male. ASR models trained on it may perform worse on female voices.
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 Chichewa-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.

CLEAR-Global/YourTTS-Chichewa: the TTS model used to generate the audio.
CLEAR-Global/XTTS-Chichewa: Chichewa XTTS model
CLEAR-Global/BibleTTS-Chichewa: related Chichewa Bible speech dataset.
Chichewa Wav2Vec-Bert ASR models: ASR models trained with different combinations of synthetic/real data

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}
}