The categorical spine
Three modes of connected speech.
Speech in the real world is not one thing. A Bantu speaker reads their own language, mixes it fluently with English, and speaks English with their own accent. We capture all three, labelled, on the same frame for every language — so a buyer can reason about the corpus before downloading it.
Mode 1 · read-aloud
Bantu only
Page-length native Bantu read aloud — gold-aligned connected speech and TTS voice.
- Gold-aligned: the read text is the reference.
- Feeds ASR, forced alignment, and a natural TTS voice.
- Page-length — carries phrasing units, not just word units.
Mode 2 · code-switch
Bantu & English
Interleaved Bantu↔English exactly as people speak — a labeled code-switch boundary in every take.
- Interleaved exactly as bilinguals speak.
- Every take carries a labeled switch point (switch_ms).
- Training data for code-switch ASR and language-ID at the boundary.
Mode 3 · accented English
English only
The same neutral English text read by Bantu L1 speakers — the accent is the asset; frontier ASR trips on it.
- Language-neutral English text — the accent is the only variable.
- Same text across speakers → a matched, comparable benchmark.
- Frontier ASR mis-transcribes it at a measurable rate.
The future row
Spontaneous narration.
Read-aloud ships today. The next row is unscripted oral narration — the same three-mode idea, but spontaneous rather than read. A translation mode (read English → speak Bantu) is reserved. The frame is built to grow.