AfriSpeech/multivoice-synthetic-speech

Synthetic Voice Samples · Africa

Synthetic speech. No human speaker was recorded for any clip in this dataset.

Generated with afrispeech-synth: text from africa-corpus, normalised to a universal orthography with africa-g2p, spoken by Google Gemini's Live API.

• —17,010 clips · 38.8 hours · 566 languages · 30 voices
• —Every clip is a distinct sentence — no sentence is repeated
• —Each language is read by up to 30 different voices, one sentence per voice
• —~1.29 hours per voice

Audio

Exactly as the model produced it — no resampling, no re-encoding, no conversion.

Two ways in

data/*.parquet carries the audio inline for training. audio/ holds the same clips as individual WAV files, which is what the browsable gallery streams and what you want if you need one clip rather than the set. The clips are identical; only the packaging differs.

Loading

from datasets import load_dataset

ds = load_dataset("AfriSpeech/multivoice-synthetic-speech", split="train")

twi = ds.filter(lambda r: r["language"] == "twi")     # one language
zephyr = ds.filter(lambda r: r["voice"] == "Zephyr")  # one voice

Rows are shuffled, so language and voice are spread across every shard: streaming the first shard gives a cross-section rather than one language in one voice.

Columns

text and normalised_text differ because universal orthography maps a language's own letters onto the shared set most African languages use — ɔ to o, ɛ to e, gy to j. The model read normalised_text.

What this is not

Recorded speech, a pronunciation reference, or evidence that a language sounds like this. A synthetic clip is a model's guess at an orthography, and quality varies enormously by language — these voices were built for widely-spoken languages and are being asked to read hundreds of others. Treat it as a starting point for bootstrapping, never as ground truth.