AfroMathVoices

African-accented English speakers reading mathematical expressions aloud, paired with target LaTeX. For speech-to-LaTeX and mathematical speech recognition research.

Hugging Face: latentic/afromathvoices

• 20 recordings of 10 prompts, read by 2 speakers (Nigerian English, Kenyan English).
• Prompts and LaTeX come from the equations_test split of marsianin500/Speech2Latex.
• Audio is mono 48 kHz Opus in WebM, not transcoded.
• Version 0.1.0 is a pilot. It is too small to train a model.

Usage

from datasets import load_dataset

ds = load_dataset("latentic/afromathvoices", split="train")
print(ds[0]["text"], ds[0]["latex"])

Files

metadata.csv   one row per recording
audio/         AMV-000001.webm ... AMV-000020.webm

Fields

One row per recording in metadata.csv:

• recording_id, file_name: recording identifier and relative audio path
• prompt_id, speaker_id: prompt identifier and pseudonymous speaker identifier
• text: sentence the speaker read
• latex: target LaTeX
• category, subcategory, difficulty: prompt classification
• duration_seconds: audio duration
• audio_format, audio_codec, channel_layout, sample_rate_hz, channels: audio properties
• source_dataset, source_split, source_sentence_id, source_license: provenance of the prompt
• dataset_version: release version

Speakers

Speaker	Accent	L1 (first language)	L2 (second language)
1	Nigerian English	Yorùbá	English
2	Kenyan English	Kiswahili and Kikuyu	English

Collection

• Speakers saw each sentence with its rendered expression in the AfriEval browser interface and read it aloud.
• Both speakers recorded all 10 prompts. They are also the dataset's authors, and each agreed to record the prompts and release the recordings under CC BY 4.0.
• LaTeX was normalized for consistent rendering in 7 of the 10 prompts, for example x^2 became x^{2} and |x| became \lvert x \rvert. The spoken text is unchanged.
• The public table omits internal task IDs, signed URLs, names, and email addresses.

Validation

Automatic checks:

• 20 records, 2 speakers, 10 prompts, 2 recordings per prompt.
• All 20 files parse as mono 48 kHz Opus.
• No two files share a checksum.
• Whisper large-v3 transcripts match the prompt text on 16 of 20 recordings after normalization. The corpus word error rate is 4.2% (8 word errors in 192 words). The differing recordings are AMV-000007, AMV-000011, AMV-000017 and AMV-000018.

Limitations

• 2 speakers cannot represent Nigerian, Kenyan, or African English.
• Two prompts (AMV-S2L-1523 and AMV-S2L-9090) have more than one natural spoken form.
• Do not use this release to compare accent performance across populations.

License

Recordings and metadata: CC BY 4.0. Source prompts are from Speech2LaTeX (CC BY 4.0); cite it when you use fields derived from it:

Korzh et al. Speech-to-LaTeX: New Models and Datasets for Converting Spoken Equations and Sentences. arXiv:2508.03542, 2025.

Citation

@dataset{amuda_irungu_2026_afromathvoices,
  author    = {Amuda, Tosin and Irungu, Chanice Wanjiku},
  title     = {AfroMathVoices: African-Accented English Mathematical Speech},
  year      = {2026},
  version   = {0.1.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/latentic/afromathvoices}
}