Words and syllables are units. Stories are the only category where a model hears language connected — a whole page, read or spoken by a native voice. We capture every language on the same three modes, so the structure is legible before you download a single file.
Page-length native Bantu read aloud — gold-aligned connected speech and TTS voice.
Interleaved Bantu↔English exactly as people speak — a labeled code-switch boundary in every take.
The same neutral English text read by Bantu L1 speakers — the accent is the asset; frontier ASR trips on it.
Mode 3 reads the same neutral English text across many Bantu L1 speakers — a matched benchmark. We ran a frontier speech-to-text model over it and scored word-error rate against the known text. It mishears Bantu-accented English at a measurable rate. That gap is the asset.
Content is held constant; accent is the only variable — so the WER is comparable across languages.
The model hears the audio and picks the wrong word — a measured accent failure mode, not silence.
The value model
A single consented Bantu–English story take is not one labelled example. Captured in one continuous take, its switch point is logged as switch_ms — so the record self-segments with no manual annotation, and one take becomes the ground truth for six things a lab would otherwise buy from six separate vendors.
Hear it proven live — a real consented take, and the datasets it yields: accent benchmark · the full 6-in-1.
License the connected-speech corpus, not a dataset.
Every language is presented on the same three-mode frame. Empty modes stay visible as the next thing to record.
138 consented takes · 98.9 min of speech · accent WER 8.2%.
54 consented takes · 116.3 min of speech · accent WER 9.0%.
1 consented takes · 2.9 min of speech.