Skip to content

Talking to Kenzy

Everything on this page works out of the box — no configuration. It's the five-minute guide to hand anyone in the house (or read aloud to a houseguest).

Waking her up

Say "Hey Kenzy." You'll hear a short chime — that's her saying I'm listening. Ask your question or give your instruction; a soft ticking sound means she's thinking, and she answers out loud.

The chime is also the privacy line: until the wake word fires, audio is processed on the room's device and never leaves it. Only after the chime does your speech travel (to your own server).

You:    Hey Kenzy
Kenzy:  ♪ (chime — listening)
You:    what's the weather tomorrow?
Kenzy:  Tomorrow looks sunny with a high of 84…

Making her stop

  • Interrupt her mid-sentence: say "Hey Kenzy" while she's talking — the wake word is the only thing she reacts to while speaking, so a bare "stop" shouted over her does nothing. "Hey Kenzy" cuts her off instantly and she listens; follow with "stop" to end it there (or say nothing and she'll go back to idle on her own).
  • The stop words — "stop", "stop it", "stop talking", "be quiet", "quiet", "hush", "shut up", "silence", "that's enough" — end the exchange silently whenever she's listening: right after the wake word ("Hey Kenzy" … "stop"), or as your answer when she's holding the mic open for a follow-up. No reply, no acknowledgment: you asked for quiet, so quiet is what you get. These are handled at the lowest level, before any thinking happens, so they're instant even if your AI model is slow or down.
  • The polite bail-outs — "never mind", "forget it", "cancel" — do the same thing but with a brief spoken "Okay, no problem," which fits the cases where you changed your mind rather than demanded silence.
  • A ringing alarm stops the moment you say the wake word — no other words needed.

“Stop” vs. “stop the timer”

A bare "stop" ends the current exchange. "Stop the timer" / "cancel the alarm" cancel the thing you named — see Timers, Alarms & Reminders.

Muting a room

"Hey Kenzy, mute" mutes that room's speaker — and it stays muted until you say "Hey Kenzy, unmute" (or "turn the sound back on"). That's the difference from the stop words above: "stop" ends one exchange; "mute" silences the room until further notice.

Two sounds intentionally survive muting, at a low volume: the wake-word chime (so you always know when she's listening) and alert chimes like the doorbell (someone at your door outranks mute). A restart also un-mutes.

Volume is voice-controlled too: "volume up", "turn it down", "set the volume to 60", "louder", "quieter".

Having a conversation

When Kenzy asks you something — a follow-up question, "should I create one called Shopping list?", the next line of a knock-knock joke — just answer. No wake word, no chime: when she stops talking in a conversation she started, she's listening. If you say nothing for about eight seconds, a soft cue plays meaning "I stopped waiting" and the conversation ends. Any bail-out phrase ("never mind") ends it politely too.

On rooms with an echo-cancelling speaker you don't even have to wait for her to finish: start talking over her during a conversation and she'll drop her volume, then yield the floor to you mid-sentence.

Conversation mode (experimental)

Kenzy can hold a real back-and-forth instead of taking one command per wake. It needs a room with an echo-cancelling speaker and the kenzy-s2s engine, and s2s.mode (Settings → Backend services) picks when it happens:

  • off (default) — always the classic one-command-per-wake.
  • on_demand — classic normally, but you can start a conversation by voice.
  • always — every wake opens a conversation.

On-demand — "start a conversation". Say "start a conversation" (or "let's talk", "continue a conversation") and just this exchange turns conversational: she answers "Okay, let's talk" and keeps listening, with a longer window that wants to stay open. When you're done, say so — "end the conversation", "stop the conversation" — and she returns to normal. That explicit close always works, even if she's misbehaving. The rest of the house stays classic.

Continue where you left off. End a conversation, and for a few minutes it stays warm: say "start a conversation" (or "continue") again and she picks up the thread — but only for you. If someone else starts one on that node, they get a fresh conversation, never your context.

Once you're in a conversation (either mode):

  • Just keep talking — reply, correct yourself ("wait, I meant the fan"), or ask the next thing, no wake word needed.
  • She listens while she thinks — change your mind mid-processing and that becomes the next turn.
  • Interrupt her freely — talking over her reply makes her stop and listen.
  • Completing a command does NOT end it — she stays available for the follow-up. Silence past the window ends it; so does the wake word.
  • Slow work goes to the background. Ask for something that takes a while — "look that up and let me know" — and she starts it, tells you so, and keeps the conversation free. If it finishes while you're still talking she comes back with the result as soon as she's done speaking; if it finishes much later she won't interrupt — she'll mention it next time you talk, announce it only if you've enabled proactive.tasks, or answer when you ask "how's that going?".

always mode turns every wake into a conversation with a shorter (~8 s) window; "turn off follow-up mode" disables it by voice (recognized voices, confirms first), from inside a conversation.

Rooms without echo cancellation, and anything that goes wrong with the engine, quietly keep the classic one-command-per-wake behavior. The conversation engine is local by default; an OpenAI Realtime cloud option exists behind an explicit opt-in — read its caveat first.

What the sounds mean

Sound Meaning
Chime after "Hey Kenzy" She's listening (plays even when muted, quietly)
Soft ticking She heard you and is working on it
Short cue after silence She stopped waiting for a reply — the conversation ended
Spoken apology Something in the pipeline failed (see Troubleshooting)
Lead-in tone before an announcement A timer or alarm you set is firing
Doorbell chime An automation rang the house (plays even when muted, quietly)
Ringing tone after "call the kitchen" The other room is being asked to accept your intercom call

Things to try

  • "What time is it?" · "What's the weather this week?" — instant, and the time works with no internet at all
  • "Turn off the lights" — your Home Assistant devices, room-aware: in the kitchen, that means the kitchen lights
  • "Add milk and eggs to the shopping list" — lists that show up on everyone's phone
  • "Set a timer for 10 minutes" · "Wake me at 7" · "Remind me to call Mom at 8"
  • "Remember that the spare key is under the blue pot" · "What do you know about the spare key?" — per-person memory with real privacy: your private facts are only ever spoken back to you (and a spoken code or password is auto-vaulted into the encrypted lockbox)
  • "Play some jazz" · "Put on The Beatles in the loft" — music by name via Music Assistant, if you run it
  • "Is Mom home?" · "Who's home?" — live from Home Assistant's presence tracking, once a person's card links their HA account
  • "Tell everyone dinner's ready" — spoken in every room at once
  • "Call the living room" — a live two-way intercom (the other side has to say "yes")
  • "Enroll me as Alice" — teach her your voice, so "my reminders" means yours
  • "Calibrate" — she tunes her hearing to your room, by voice

The full list of built-in abilities is in Built-in Skills.