the hotword saga. the assistant app builds a persistent overlay when the screen is on, minimal, darker, more opaque, just an open app icon and a close icon, but the whole point was saying the wake phrase with earbuds in and having a conversation start. it didn't fire.
the debug view told a confusing story. mic levels were healthy, the chunk counter kept climbing, levels crossed way past halfway. the audio pipeline was fine. detection was the dead end. first real bug, the audio model outputs a four dimensional tensor and the code expected three. fixed the cast and rebuilt the wake test. still, the trigger level sat at zero while i stood there saying the phrase over and over.
so we retrained the wake word with six real recordings of my voice. round two over-augmented those clips into hundreds of synthetic variants, and the model got worse, not better. real recordings beat synthetic copies, turns out. restored the earlier recipe, a pool of a couple hundred positives including the real clips, and validated against clips the model never saw. three of the six real clips fired at high scores, the synthetics stayed strong, and zero false triggers.
the three that didn't fire probably weren't clean reps. the move now is more real recordings, said in a normal steady voice a few seconds apart, two passes, enough to train and validate properly on held out samples. then ship the build with a live score line so the trigger is visible in real time instead of buried in logs.