When a feature 'does nothing' in a running app, read its live settings first: a mute flag is cheaper to find than a pipeline bug and can be flipped over the app's control channel
A user said saying the trigger word produced no output. Before reading detector code, read the app's live settings (here the dashboard WebSocket sends a 'settings' message on connect: name_alerts, find_cues, sound_alerts, names...). name_alerts was 'off', which by design means 'still logged, no buzz', so the whole recognition chain could be fine while the user perceives nothing. Fix was one message over the same WebSocket the UI uses ({cmd:'setting', setting:'name_alerts', value:'on'}), no restart. Also note a state-sync trap: a separate command that changes the live listener ({cmd:'name', name:'Max'}) did NOT update the settings object the dashboard displays (settings.names stayed 'Shravan' until the 'setting' command was used), so the UI and the engine disagreed; prefer the command that goes through the settings layer, and verify by re-reading the settings echo. Runtime changes are lost on restart if the same value is also a CLI flag (--name), so tell the user to change the launch command too. A 40-line aiohttp ws_connect script that sends one JSON command and prints 'log'/'settings' replies is enough to drive and verify this from the terminal. Tip for the interaction itself: spoken-keyword detectors usually need a cooldown (here 1.5 s) and ~1 s of transcription latency, so test by repeating the word with a pause rather than rapid-fire.