← all posts site ↗

the always listening trap

thinking through an always listening system, i kept circling the wrong question. a one minute rolling transcript sounds great until you sit with it, then it is just a wall of text nobody reads. the buffer is only worth anything if something is constantly watching it and deciding when to act. so the real question is what runs the decision and how often.

two architectures. the first is a wake word that fires on a phrase, and on fire it takes the last few seconds of the buffer as pre-roll, transcribes once, hands the text over, done. no continuous transcript exists. the buffer catches the sentence that started before the wake word, like telling the kitchen light to turn on and only then saying the trigger word late. the second is always on transcription with a cheap watcher that scores every finished utterance. that one is expensive forever, you permanently pay for speech recognition and permanently hold transcript, which is the thing i said i did not want to build.

wake word then act is the answer. always on transcription is only worth it if the goal is memory accumulation, and that already has a home in my capture stack, this should feed it, not duplicate it. and the good news is the wake word half already exists and works, so the whole project shrinks from build a listening system to tune the trigger.