Afri2Universal: Universalized Multilingual African Speech Corpus (314 Languages)

Afri2Universal is a large-scale, phonetically universalized speech recognition dataset 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 26-character universal latin script (lowercase a-z, stripped of punctuation and diacritics) designed for acoustic-phonetic cross-lingual transfer learning, low-resource speech recognition, and parameter-efficient LoRA fine-tuning on foundation speech encoders (such as Meta's omniASR / wav2vec2 300M).

---

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: ~950.7 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: 292,791 utterances (~950.7 hours)
  • dev: 15,571 utterances (~50.0 hours; strictly held-out 5% partition)
• Vocabulary: 30 tokens (4 special tokens <pad>, <unk>, <s>, </s> + 26 latin letters a-z)

---

Data Schema

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

Column	Type	Description
text	string	Universalized normalized 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)

---

Vocabulary and Tokenizer

The repository includes a standalone SentencePiece character model in vocab-univ/:

• vocab-univ/char_spm.model: Character tokenizer model (vocab size: 30)
• vocab-univ/char_spm.vocab: Vocabulary token list
• vocab-univ/chars.json: Character alphabet mapping (a through z)
• vocab-univ/language_distribution.tsv: Per-language utterance and duration distribution

---

Fairseq2 & OmniASR Integration

The cards/ directory provides pre-configured asset cards for seamless training with fairseq2:

# cards/afripa_univ.yaml
name: afripa_univ_tokenizer
tokenizer_family: char_tokenizer
tokenizer: /data/vocab-univ/char_spm.model

---

name: afripa_univ_speech
dataset_family: mixture_parquet_asr_dataset
dataset_config:
  data: /data/corpus-univ/version=0

---

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:", sample["transcript"])
print("Universalized a-z 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

If you use this dataset, please cite:

@dataset{afri_uni_asr_2026,
  author    = {AfriSpeech and Ghana NLP},
  title     = {Afri2Universal: Universalized Multilingual African Speech Corpus Across 314 Languages},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AfriSpeech/afri2universal}
}