One recording does the work of six datasets

In a tonal Bantu language the meaning selects the tone. A native speaker needs only the intended meaning to say a homograph correctly — so the recorder shows the Bantu word and, beneath it, a “means … (English)” prompt. Nothing else.

That one prompt does triple duty — it disambiguates for the speaker, it is the English translation label, and it is the English half of the code-switch (after switch_ms). So a single consented take self-labels into six datasets. The word is recorded across 5 rounds — reproducibility you can measure and train on.

Live example: 🇿🇲 Bemba 🇿🇦 Xhosa 🇸🇿 Swati 🇰🇪 Swahili 🇷🇼 Kinyarwanda your access: public

The speaker only needs the meaning. Shown the word + “means …”, they produce the correct tone — and that single take self-labels into six datasets a lab would otherwise buy from six vendors.

Iintanga

means Age mates

Bantu 0–5145 ms · English 5145–11280 ms · FSI: i·i·nta·nga

live

DATASET 1

Bantu ASR

Bantu audio (0→switch) + its transcript.

live

DATASET 2

Code-switch ASR

The whole utterance — Bantu + “means …” — the rare code-switch ASR.

live

DATASET 3

Bantu↔English pair

Aligned Bantu↔English parallel pair.

live

DATASET 4

TTS / voice

Clean 48 kHz consented voice + text.

live

DATASET 5

Code-switch boundary

The exact language boundary (switch_ms) — no manual annotation.

roadmap

DATASET 6

Tone / F0

Pitch/F0 over the syllables — the tone the spelling hides.

Tone/F0 activates per language as its forced-alignment model ships; the other five are usable today.

Showing Iintanga — “Age mates”. Audio spans play at evaluation tier and above.

Start an evaluation → play every span, every round

See what we have & what you can use today · the full corpus (all rounds, all languages, bulk) is licensed.