Contributors state where they are from and read phrases from a 320-item bank in the voice they use at home; the brief says not to neutralize, over-enunciate or imitate an accent that is not theirs, and flattened or performed recordings are rejected. Each clip carries its self-reported origin and prompt, a weak accent label and reference text at no annotation cost, and the shared bank puts the same phrase in many accents; ten clips per speaker at most. Accent bias in ASR is well documented and more of the standard accent does not fix it; the origin tag is the label a scraped corpus cannot give you, so under-represented accents can be selected.