The wake word

Saying the assistant’s name should start a turn. Everything else you say in the room should not.

That sounds like a small problem until you try to solve it by reading text.

Why not transcribe first

The obvious approach is to transcribe everything and check whether the words came out right. It is obvious, and it is what this used to do, and it does not work.

A speech engine has to guess at a name it has never seen. Ask it to hear Jane and it may write Jane, or Jayne, or Jain, or “jay n”. None of those are mistakes exactly — they are reasonable spellings of a sound — but a matcher comparing letters sees three different words. You can loosen the comparison until the variants pass, and then unrelated speech starts passing too. Tighten it and your own name stops working. There is no setting that is right, because the problem is not in the setting.

Worse, transcribing first means everything gets transcribed. Something you said to another person in the room has already been through a speech engine before anything decides it was not addressed to the assistant.

What it does instead

The gate runs a speech recogniser that can only produce one phrase. Your wake words are compiled into the sounds that make them up, and the recogniser is restricted to that sequence and nothing else. Ordinary speech does not come out wrong — it does not come out at all.

Nothing is spelled, so nothing can be misspelled. And because the decision is made from sound, it happens before transcription: an utterance that is not the wake phrase never reaches a speech engine at all.

The model is small — a few megabytes, running on the processor rather than the graphics card — and it is open source under a permissive licence.

It tells you while you are still talking

The recogniser reads your speech as it arrives, so the phrase registers the moment you say it — not when you stop. A green heard you marker appears mid-sentence, and you can keep going knowing the rest is being taken.

This matters more than it sounds. Deciding at the end meant you could say the phrase, talk for a minute, and only then find out nothing had been listening — with no way to tell earlier, because nothing had decided yet. If the phrase does not land, you know within a breath rather than a minute.

It needs no training

This is the part worth being clear about, because most wake words do not work this way. Alexa and Siri cannot let you choose a name, because a custom phrase would mean training a new model on thousands of hours of audio. That is why their names are fixed.

Here, your phrase is turned into sounds from a vocabulary the model already knows. It takes milliseconds and it happens again automatically whenever you change the name or the prefix. Any two words, no training, no waiting.

Checking it hears you

What the guided run in Settings does is not training — it is checking, and finding the right sensitivity for your voice in your room:

1. Say the phrase five times.
2. Read three short lines that do not contain it.

The first set has to be heard every time. The second set has to be heard never. The app tries the tightest setting first and works down only as far as it must, so you get the strictest sensitivity that actually hears you rather than the loosest one that happens to work.

If it cannot find one, nothing is saved. You are told which half failed: your takes were not heard, or your ordinary speech was. Those are different problems with different fixes, and a setting stored from a run that failed either way is a wake word that misbehaves while you have no reason to doubt it.

You can run it as many times as you like. It is worth redoing if you move rooms, change microphone, or find it firing when it should not.

Choose a distinctive name

If your ordinary speech keeps firing the gate, the name is usually the problem, not the setting.

A name that is also an ordinary English word appears in ordinary sentences. If the assistant is called Assistant, then “the assistant said she would call back later” contains the phrase — because it partly is the phrase. No sensitivity separates them, and the honest answer is to change the name rather than to widen the gate until conversation starts turns.

Nothing leaves the machine

The recordings you make during the check are decoded in memory and dropped. They are not written to disk, not uploaded, and not kept. What is stored is your phrase and two numbers.

There is nothing there to leak, and the whole thing keeps working with no network at all.

And it does not listen through a mute. When the microphone is off, it is off — there is no low-power always-on path waiting behind it. The design could support one and deliberately does not: a mute that is not a mute is a promise you cannot take back.

Three states, and they are different

	what happens
Off	Every utterance is dispatched.
On, not yet checked	Every utterance is dispatched.
On and checked	Only the phrase wakes it.

The middle row is deliberate. Turning the switch on is permission; the check is readiness. Until you have said the phrase and watched it land, the gate stays open rather than going deaf — an assistant that hears nothing is far worse than one that hears everything, and a sensitivity that suits one voice may not suit yours.

Everything that can go wrong lands in that same open state. A missing model, a setting file it cannot read, audio too short to decide on, a renamed assistant whose new phrase was never checked, even a name the model has no sounds for — all of them dispatch the utterance and say so. Only a confident miss discards.

That includes the moments just after the app starts, and just after you finish the check. Loading the recogniser takes several seconds, and until it is ready nothing has heard your speech at all — so those utterances go through rather than being refused by something that was not listening yet. Nothing waits on it either; the load happens in the background while the app stays responsive.

Renaming the assistant clears the check, because what was confirmed is that two particular words are heard reliably, and the new name is not among them.

Where it lives

Settings → Voice runs the check, and shows which of the three states you are in. Forget clears it and returns the gate to hearing everything — that is the whole undo, and it needs no config edit.

Two settings live in identity.wake_word in mirror.yaml:

• min_threshold — a floor you can raise above what the check found, if you want the gate stricter than your recordings implied. It cannot lower the bar below what was confirmed.
• boost — how strongly the recogniser favours your phrase over what it would otherwise hear. The shipped value is neutral — it favours nothing. Raising it makes a mumbled phrase more likely to land, and false wakes more likely with it.

Neither is a speaker check. Anyone who says the phrase wakes it.