Each clip is a single phrase from a 2,800-item bank, read in the contributor's natural voice on their own phone with a beat of silence before and after. Clipped ends, wrong phrases and drowned-out readings are sent back; artificial accents and text-to-speech are excluded. Every clip maps to its item in the bank, so reference text comes without a transcription pass, and the same phrase is spoken by dozens of different people, useful for accent contrast and TTS work. Speakers come from a global base, Southeast Asian, American and European English foremost, on the devices and in the rooms a deployed model meets.