Afri2Universal: Universalized Multilingual African Speech Corpus (314 Languages)

Afri2Universal is a large-scale, phonetically regularized speech recognition corpus spanning 314 African languages and dialects, totaling over 950 hours of audio balanced across 4 major African language families (Niger-Congo, Afroasiatic, Nilo-Saharan, and Khoisan).

The dataset is orthographically regularized into a universal 26-character Latin representation (lowercase a-z, stripped of punctuation, with natural word boundaries preserved) using Africa-G2P. This enables cross-lingual acoustic transfer learning, parameter-efficient fine-tuning on foundation speech encoders (Whisper, omniASR), and robust speech recognition across low-resource African languages.

---

Orthographic Regularization via Africa-G2P

All universal target texts in this dataset were generated using Africa-G2P (africa_g2p), an open-source rule-based grapheme-to-phoneme and orthographic normalization toolkit developed for African languages.

• Universal Mode (UNIVERSAL): We utilized GraphemeConverter(lang, UNIVERSAL) to systematically map language-specific characters, diacritics, and tone marks to standard Latin equivalents (e.g. ɔ → o, ɛ → e, ɗ → d, ɓ → b, and stripping tone accents).
• Non-Latin Transliteration: For non-Latin scripts (such as Ge'ez for Amharic/Tigrinya and Arabic script), africa-g2p provides consistent phonetic transliteration into the universal Latin character set.
• Word Boundaries Preserved: Unlike raw character streams, single whitespaces separating words are preserved, allowing standard speech models (such as Whisper) to learn real word boundaries and true Word Error Rate (WER) evaluation.

Example Conversion:

• Original Transcript: Onyankopɔn, me din de Kwame! Ɔyɛ kɛse.
• Africa-G2P Universal Target: onyankopon me din de kwame oye kese

---

Dataset Summary

• Total Languages: 314 African languages / dialects (all ISO 639-3 tagged)
• Subsets / Configurations: Select all for the full multilingual pool or select any individual language code (e.g. twi, yor, hau, swh, ibo, wol, lin, som, amh, etc.) from the subset dropdown above!
• Total Audio Duration: ~950.6 hours (train) + ~50.0 hours (held-out dev)
• Target Duration: Balanced to ~8.39 hours per language partition
• Sampling Rate: 16,000 Hz (FLAC audio)
• Splits:
  • train: 292,591 utterances (~950.6 hours)
  • dev: 15,563 utterances (~50.0 hours; strictly held-out 5% partition)
• Vocabulary: 26 lowercase Latin letters (a-z) + space

---

Data Schema

The dataset is partitioned by split and language in Apache Parquet format:

Column	Type	Description
text	string	Africa-G2P universalized target text ([a-z ]+)
transcript	string	Original surface orthography of the utterance
audio_bytes	audio	16 kHz mono FLAC audio (playable in the viewer above)
audio_size	int64	Total number of audio samples (duration * 16000)
duration	float32	Duration in seconds
ipa	string	Phonetic placeholder (empty in universalized target)

---

Usage

Loading a Specific Language Subset

from datasets import load_dataset

# Load a specific language partition (e.g. Twi)
dataset = load_dataset("AfriSpeech/afri2universal", "twi")

# Access train or dev splits
sample = dataset["train"][0]
print("Transcript (Original):", sample["transcript"])
print("Africa-G2P Target:", sample["text"])
print("Duration (s):", sample["duration"])

# Audio is decoded automatically as an Audio feature
audio = sample["audio_bytes"]
print("Audio array shape:", audio["array"].shape)
print("Sample rate:", audio["sampling_rate"])

Loading All 314 Languages

from datasets import load_dataset

# Stream all languages
dataset = load_dataset("AfriSpeech/afri2universal", "all", streaming=True)
for sample in dataset["train"]:
    print(sample["transcript"], "->", sample["text"])
    break

---

Citation & Acknowledgements

This dataset relies on the Africa-G2P library developed by AfriSpeech for phonetic and orthographic regularization.

@software{africa_g2p_2024,
  author       = {AfriSpeech},
  title        = {Africa-G2P: Grapheme-to-Phoneme and Universal Orthographic Converter for African Languages},
  year         = {2024},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/AfriSpeech/africa-g2p}}
}

@dataset{afri2universal_2026,
  author    = {AfriSpeech},
  title     = {Afri2Universal: Universalized Multilingual African Speech Corpus Across 314 Languages},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AfriSpeech/afri2universal}
}