Nine real failure modes, in the order they happen — the click that does nothing, the microphone another tab is holding, the silence that is deliberate, and the edit she says she made that you cannot see.
You clicked the little face in the corner, said “can you move the dishwasher one over,” and got nothing back. No voice, no movement, no error. That silence has about six different causes, and only one of them is a fault — the rest are a permission dialogue hiding behind another window, a microphone another tab grabbed first, or a deliberate standby that closed the line while you were thinking. This page is the honest list, in the order the problems actually happen, plus the two things people ask before they trust a microphone at all: what the free time is, and where the audio goes.
Emma sits in the corner of the planner canvas, in the 2D plan and in 3D. Left-click starts a voice call. Click her again to hang up. Right-click opens the typing panel. That is the entire control surface — there is no settings drawer, no push-to-talk key, no separate microphone button somewhere else on the screen.
What most people miss is that how solid she looks is information, on purpose. Faded means she is not listening — the microphone is not open, nothing is being sent, and nothing is being spent. Solid, with a ring around her, means the line is live. If you are ever unsure whether you are being heard, do not judge by whether she answered. Look at the corner.
She also fades further while you hold the mouse button down on the canvas. That is deliberate: the seconds when you are actually dragging a cabinet into position are the seconds she should be out of your way. Release the mouse and she returns to whatever state she was already in. It is not a dropped call and it is not a bug.
Four things have to be true before a voice call can start — a browser that supports live audio, a Google sign-in, microphone permission granted to this site, and time available on the account. If any one of the four is missing, the click does not fail silently. It drops you into the typing panel and tells you which of the four was the problem. Read that line before you read the rest of this page; it is the fastest diagnosis here.
What they search at 11pm “AI voice assistant not working” — usually while the microphone is perfectly fine and the browser is asking a permission question behind another window.
These feel identical and are not. If clicking her opens the typing panel instead of starting a call, the fallback is doing its job and the reason is written in the panel. Two causes account for almost all of it: you are not signed in, or the browser you are using has no support for live audio. Sign in with Google, or open the planner in a current version of Chrome, Edge, Safari or Firefox, and click her again.
The other case is worse to sit through: the call starts, the ring appears, and she never says anything back. That means the connection is up and the fault is upstream of her, in the microphone. Work through these in this order, because they are ordered by how often they turn out to be the answer.
Try a private window with extensions disabled. Privacy extensions and ad blockers occasionally block the connection that carries live audio, and the symptom is a call that appears to start and then does nothing. If a private window works, an extension is the culprit and you can find it by re-enabling them one at a time.
And the honest limit: some managed work computers block microphone access to all websites at the policy level. Nothing we ship can override that, and no amount of clicking Allow will help. On those machines, right-click her and type instead.
This is the report we get most, and nearly every time it is standby working exactly as designed. After five quiet minutes, the connection closes on purpose. You did not lose your internet and she did not crash. The line was open, nobody was talking, and holding it open would have cost you time you were not using.
Three things stay true while she is in standby, and they are the parts worth knowing:
There is a short beat between you starting to talk and her being back on the line. Talk through it rather than stopping to check — the first second or so of your sentence is what wakes her, and if you keep going she catches the rest of it.
The reason standby exists at all is arithmetic, not cleverness. Live voice is paid for in five-minute blocks. A line left open through a long silence spends a block on an empty room, so the line closes and the microphone waits instead. It is the same instinct as switching off a light in a room you left.
If the ring is still showing and she does not answer after you have spoken a complete sentence at normal volume, that is not standby — standby has no ring. That is a microphone fault, and it belongs to the list in section 01.
This one is genuinely confusing, because she is not lying and you are not imagining it. There are four causes, and the fourth is responsible for more head-scratching than the other three put together.
If she moved the wrong thing, say so immediately, while the numbering still means what it meant when you spoke. Waiting until you have made three more changes turns a one-sentence correction into an archaeology problem. For the full list of what she can and cannot change, see the voice commands page; for how she knows what is on your screen in the first place, that is explained here.
She is instructed to tell you why she is not doing something rather than quietly doing something adjacent and hoping you do not notice. That makes a refusal genuinely useful information about your kitchen, and it is worth reading it as a finding rather than a failure. Four of them cover almost everything.
| What she says | What it actually means | What to say next |
|---|---|---|
| There is no room on that side | The gap left between the neighbouring cabinet and the wall is narrower than the narrowest cabinet that exists in that family. Real cabinets come in real widths. | “What will fit in that gap?” or ask her to narrow or remove the neighbour first, then place it. |
| I cannot find anything numbered 7 | The list has changed since you last heard it — almost always a deletion that renumbered the items after it. | “List the cabinets again,” then use the number she gives you now. |
| You need to be signed in for that | The tools that change the design need an account behind them. | Sign in, then click her again to start a fresh call. The design in front of you does not go anywhere. |
| That would push it into its neighbour | The move collides with something fixed. She will not overlap two cabinets to satisfy an instruction. | Give her a distance instead of “over a bit”, or ask her to shift the whole run rather than one box. |
The first refusal is the one people argue with, and it is usually right. A wall is whatever length the builder made it; cabinets come in fixed widths; the leftover is what fillers are for. If the refusal is about the kitchen rather than the software — a run that will not close, a corner that will not work, an appliance that will not fit — then it is a design problem and the chapter on fixing things is a better read than another attempt at phrasing.
To interrupt her, just start talking. She stops. There is no key to hold and no polite gap to wait for, and you should use it far more than feels natural — if she is three sentences into explaining something you already know, talk over her. That is how the conversation is meant to go.
To type instead, right-click the avatar. The typing panel will answer questions and explain what it can see, but it has no tools: it cannot move, add or delete anything. Only a voice call can edit the design. That split is deliberate rather than an oversight, and the practical consequence is that a typed “move it left” earns you a helpful explanation of how to move it left, not a cabinet that has moved.
She works in English, Mandarin Chinese and French. Speak the language you want to be answered in.
And then the part nobody in software likes writing down: voice is not the fastest way to do everything, and using it for the wrong jobs is the most common reason people conclude it does not work.
Reach for the mouse instead when the job is any of these:
If your problem is that she misunderstands what you meant rather than that she cannot hear you, the phrasing that works is collected in what to say to an AI kitchen designer.
Live voice runs in five-minute blocks. The block is the unit — it is not metered by the second, and that is why standby matters so much.
Every account gets ten free minutes a day, which is two blocks. The day rolls over on the Toronto calendar day, so if you are designing from another time zone, that is the clock to picture rather than your own midnight. After the free time is used, voice costs 1 GH Coin per five-minute block. Typing to her costs nothing at all, and has no blocks.
The rule that keeps this fair is the renewal condition: a block only renews if you actually spoke during it, and spoke within the last minute and a half. A call you walked away from does not quietly buy another block behind your back. Put that together with standby closing the connection after five quiet minutes, and it is genuinely difficult to spend time on an empty room.
| What you are doing | What it uses |
|---|---|
| Typing to her | Nothing. No account needed, no blocks. |
| Your first ten voice minutes each day | Free — two blocks, reset on the Toronto day |
| Voice after the free time | 1 GH Coin per five-minute block |
| Sitting in standby | Nothing. The connection is closed; only your microphone is idling. |
| A call nobody is speaking in | Nothing more — the block will not renew without speech. |
When you see the message that the free time is used up, it means exactly that: the daily allowance is spent and there is no coin balance to carry the next block. Nothing else has broken, and your design is untouched. You have three options, and the third one is free forever.
Everything else in the planner — drawing the room, placing cabinets, countertops, colours, the 3D view, the quote — is free and has nothing to do with these blocks. The clock only runs while a voice call is live.
People ask this before they let a website open a microphone, and they are right to. Here is the whole of it, plainly.
Talking to Emma by voice is not recorded at all. The audio goes from your browser straight to the speech provider that turns sound into words and words back into sound. It does not reach us. We do not keep a transcript of it. When the call ends there is nothing on our side to replay, because there was never anything to replay in the first place.
Typed conversations are kept — what was asked and what was answered. That is the honest difference between the two panels, and it is worth knowing when you choose which one to use. If you would rather a particular question were not stored anywhere, ask it out loud.
During standby the microphone stays on but nothing is transmitted. The page measures the loudness of the room locally, purely so it can notice that a person has started speaking. It is not listening for words while it does that, and it is not sending audio anywhere. That is also why she wakes to a spoken sentence and not to a cough: a level is all the page has to work with.
All of this is written down in our privacy policy, which is the version that governs rather than this paragraph.
The limit of what we can promise: we can tell you what our own software does, and we have. We cannot make promises on behalf of your browser, your operating system, or an extension somebody installed on your machine. If there is something confidential being said in the room, end the call rather than trusting a status light — not because we know of a problem, but because that is sound practice with every microphone on every site you will ever use.
Four things have to be true before a voice call can begin: a browser that supports live audio, a Google sign-in, microphone permission for the site, and time available on the account. If any one is missing, clicking the avatar opens the typing panel instead and tells you which of the four was the problem. Read that message first, because it is the diagnosis.
Check in this order. First, look at the avatar: solid with a ring means the call is running and the microphone is the suspect, faded means no call started. Second, open the padlock menu in the address bar and confirm microphone access is allowed, and that it is pointed at the input you are actually speaking into. Third, quit any other application or tab holding the microphone, such as a video call, because the first application to claim it usually keeps it. Fourth, check the default input in your operating system's sound settings. Reload the page after any change, because permission is read when the call starts.
Almost always standby, which is deliberate. After five quiet minutes the connection closes so an open line does not spend your time on an empty room. Your microphone stays on but nothing leaves your computer, because the page only measures the loudness of the room locally. Speak for a moment and she comes back with the conversation intact. After about twenty minutes of silence she releases the microphone entirely and the call is over.
No. Voice conversations are not recorded at all. The audio travels from your browser straight to the speech provider and never reaches us, and we do not keep a transcript. Typed conversations are different: what was asked and what was answered is kept. Both are described in our privacy policy.
Every account gets ten free minutes a day, which is two of the five-minute blocks voice runs in, and the allowance resets on the Toronto calendar day. After that, voice costs 1 GH Coin per five-minute block. A block only renews if you actually spoke during it and spoke within the last minute and a half, so a call you walked away from does not quietly start another one.
For voice, yes. A live call needs a Google sign-in, because the tools that change your design need an account behind them. The typing panel needs no account: right-click the avatar and ask away. Drawing the room, placing cabinets and viewing it in 3D are free either way.
Four common causes. You were in the 3D view and editing had to be switched on first, so the change came a beat after the view changed. Several items were selected, so the planner treated them as a group. The item is inside a packed group and has to be unpacked before it can be addressed on its own. Or the number moved: every removal renumbers everything after it, so the number you agreed on a minute ago now points at something else. After any deletion, ask her to list the items again.
Yes. Start talking and she stops. There is no key to press and no gap to wait for, and interrupting is the intended way to use it rather than a rude edge case.
Right-click the avatar to open the typing panel. It is free, needs no microphone, and will answer questions and explain what it can see. What it cannot do is change anything: the typing panel has no tools, so only a voice call can move, add or delete cabinets. She works in English, Mandarin Chinese and French.
Open your plan, left-click the face in the corner, and ask her what is wrong with the wall you are least sure about. Ten minutes a day are on us.
Open the 3D planner →