AfriSpeech/african-speech-ipa

African Speech IPA

AfriSpeech audio paired with IPA phoneme transcriptions, for 141 languages.

Each row has the audio, the source transcript, and ipa — the transcript converted to space-separated IPA phoneme units with africa-g2p 0.2.0. Units are kept whole, so t͡ʃ, k͡p and kʰ are one token each rather than two or three characters, and punctuation is preserved as its own unit so alignment and TTS keep their phrasing.

from datasets import load_dataset
ds = load_dataset("AfriSpeech/african-speech-ipa", "yoruba_yor", split="train")
ds[0]["ipa"]          # 'ʃ e k ú n'

How each language was phonemised

africa-g2p carries rules for 400 languages, but not for every language here. Each is mapped to a rule set in one of three tiers, and language_map.json records the choice and its measured character coverage for all of them.

Donors are chosen by family first, then geography, and only kept if they resolve at least 95% of the language's characters. Family is a hard filter rather than a ranking signal: coverage alone is blind to whether a chart should be used, and it happily scored a Germanic donor at 1.0000 for a Bantu language. One language (Damara) is excluded, because its clicks are spelled with IPA click letters no available rule set is related to.

Coverage is measured on these transcripts, not on a parallel corpus — scoring Hausa against one put it at 0.74, because that corpus mixes Ajami with Boko and no Latin chart resolves Arabic script.

Files

• —data/<language>/{train,validation,test}-*.parquet — the corpus. Splits are AfriSpeech's own, which are speaker-disjoint; a held-out slice of train would not be.
• —language_map.json — rule set, tier and coverage per language
• —vocab/units.json — the phoneme inventory, most frequent first
• —vocab/proxy.json — unit to Private Use Area codepoint. Only needed to reproduce CTC training against a character tokenizer, which would otherwise split multi-character units.
• —vocab/counts.json, vocab/per_language.json — unit frequencies, overall and per language

Code

AfriSpeech/afrispeech-ipa-trainer builds this dataset and trains on it. Phonemisation rules are africa-g2p.

Caveats

• —Phonemisation is rule-based, so it reflects the orthography rather than a speaker's actual pronunciation. It does not model reduction, assimilation or dialect.
• —Donor-tier languages borrow another language's letter-to-sound rules. Coverage says the chart can spell the words, not that every value is right for that language.
• —Amharic, Kpelle and Loma are written in syllabic scripts, so their units are CV syllables rather than single phonemes.
• —Clips are 0.4–39 s. Anything whose phoneme count exceeded what CTC could align against the audio was dropped.
• —Punctuation is kept as its own unit, but not all of it is speech. Sentence marks (. , ? ! ; : and quotes) reflect something audible — a pause, a contour — while brackets and editorial markup (( ) [ ] * ¶ § •) come from the written source and have no acoustic correlate at all. vocab/counts.json lists every mark with its frequency; filter ipa on the ones you want rather than assuming the set is uniform.

Per-language detail