the wake word kept firing on tv shows and youtube videos, so the fix was to make the wake gate care about who said the phrase, not just that the phrase was said. the app already had a voiceprint model, the same one the lifelog voiceprints use, but it had been enrolled on only three samples. re-enrolled it properly with fifty six real takes, the wake clips plus the calibration recordings, and the separation came out clean. my takes sit at a median cosine of 0.66, every other voice tops out around 0.11, and the gate sits at 0.30.
shipped it as a toggle with a one time model download. after that, any trigger runs through the voiceprint first, so the tv, a youtube video, and a podcast saying the wake phrase all get ignored, and the log shows the match score on every attempt. the gate fails open, which is the important part. if anything is missing or broken, the wake word behaves exactly like before, so detection can never brick itself.
phase two is voice id on interrupts, which needs raw mic audio during a session, and the speech recognizer currently owns the mic. the next step is finding out whether the phone lets a second listener read the mic alongside it. the wake side alone already kills the biggest annoyance though, which was the tv starting conversations.