The proof point

Frontier ASR trips on the accent.

Mode 3 has many Bantu L1 speakers read the same language-neutral English stories. We transcribed them with a frontier speech-to-text model (google-cloud-stt en-US/latest_long) and scored word-error rate against the known text. The errors are mostly substitutions — the model hears the audio and picks the wrong word. That measurable gap is the asset.

Accent (L1)	Takes scored	Reference words	Word error rate	Word accuracy
Bemba	46	27328	6.8%	93.2%
Luganda	40	23150	15.1%	84.9%
LUN	25	14906	17.5%	82.5%

Hear it

A public sample take.

Visiting Grandmother in the Village · Bemba-accented English · 675.3s · 759 words · ASR WER 10.4%

0–2640 ms The Morning at Home

2640–5400 ms I woke up early.

5400–8640 ms The house was quiet.

8640–12000 ms The sun was not yet high.

12000–15480 ms I opened my eyes slowly.

15480–18660 ms I listened for a moment.

18660–21960 ms I could hear a bird outside.

21960–25620 ms I could hear someone sweeping the yard.

Why it's a clean benchmark

Same text, different mouths.

Because every speaker reads the same neutral English stories, content is held constant and accent is the only variable. That makes the WER comparable across languages and directly usable as a robustness benchmark for any English ASR system.

Coverage matrix Request the eval set

Frontier ASR trips on the accent.

A public sample take.

Same text, different mouths.

Start where you are.