The problem
A consumer device team had a wake word that worked in Delhi and failed across most of eastern and southern India. Their existing corpus had been collected in three metros from volunteers on recent phones.
What we did
We recruited against district-level quotas rather than state-level, which is where the accent variation actually lives, and required each speaker to record at three distances in two room types.
Capture-time checks rejected clipped and silent recordings on the device, so the review queue only ever saw usable audio.
Outcome
False rejection on the previously weak regions dropped substantially, and the team could point at which districts were still under-represented rather than guessing.
