wake word training turned into a full accounting problem. i'd sent ten takes from the phone and only eight made it into the retrain. one ten second take from the night before had landed in a folder the trainer never reads, so i had it copied back into the input pool. three of its five clips fire hard on the current model, so it was real signal sitting in a dead folder.
two takes were junk by design. one was an emulator test that came out silent, and one was pure silence, peak level eight out of thirty two thousand, a mic glitch where nothing ever got captured. not a pipeline failure, just a dead recording.
the weirder finding is that the retrains kept producing worse models than what's on the phone. both candidates scored below the current model, so the safety gate rejected them and kept the old one. the recordings are fine, the recipe is the weak link.
new plan: record one long take and say the wake word over and over, thirty reps, then another take with earbuds in, another from across the room, loud and quiet, nothing expires, and all of it trains together. the splitting happens at training time, not while recording. more varied samples, or the next retrain stays a coin flip.