Identity switches need to be hearable

What it is

A living-room speaker moves among accounts: the last utterance was on the father’s calendar, the next is already in the mother’s shopping list. On a screen, an account switch has an avatar. On the voice channel, without a hearable announcement, people in the room keep speaking as the previous identity. The announcement is for the whole room, not a line of type for the mouth that was recognized. What is required is that the switch itself be heard, even when recognition was correct.

Why it happens

Speech in a room is a shared channel. Identity is a private index at account grain: calendars, contacts, recommendations, home permissions all differ. A GUI pins “who it is now” in a corner of space. Audition has no corner; the previous turn’s identity stays in working memory until a new form of address overwrites it. If the system switches silently, later commands hit another store while the speaker still thinks they are operating their own. People who did not trigger the switch hear even less — they may ask “has my parcel arrived,” with “my” still bound to the last identity. The announcement has to occupy auditory bandwidth to count: name the person (“switched to Wang’s account”), name the object about to move (“added to Wang’s list”), so a wrong identity can still be stopped before execution. Flashing an avatar in a phone companion app does not tell the room.

Studying it

Run silent switch against spoken naming. At least two people in the room; one triggers the switch, the other then issues an identity-bound command (calendar, list, message). Dependent measures: can the person who did not trigger it say whose data would change, and how often the wrong account is altered. Do not stop at satisfaction of the person who was recognized — the cost of the shared channel lands on the person who did not speak.

In logs, mark “account changed, no name spoken this turn” as a missing announcement, then see whether the next three turns touch calendar, messaging, or payment — objects bound to identity.

Where it stops holding

A home with one enrolled user never switches; there is no one to announce. In a car with only the driver, spoken naming may be surplus — until a back-seat passenger can wake the system, at which point the shared channel returns. When the user says “use my mother’s account,” the command is already an announcement; repeating it still helps as confirmation of the execution target, and can be shorter. Reading a full account bio on every wake turns announcement into noise that people learn to ignore.

Applying it

• When the account changes relative to the previous turn, put a name or role in the first reply: “this is Mum’s calendar.” Do not leave the name only in a phone push.
• When about to rewrite an identity-bound object (list, event, message, home permission), speak whose object it is. Shared objects (the living-room lamp) need no naming.
• Offer a spoken correction: “not me — switch to Dad.” The correction itself must be hearable in the room.
• How to check: two people take turns calling the same device; a third sits in earshot and after each turn is asked “whose thing just changed.” Turns they cannot answer are turns the announcement did not enter the ear.

Related

• Same group: M4.08.1 Mismatch between enrollment and use environments lowers hits · M4.08.3 Uncertain recognition should fall to the least-privileged identity
• Nearby: M4.03 Multi-User Recognition · M1.08 Turns and Floor Management · M2.10 Persona and Consistency
• Search terms: audible identity switch · account announcement · shared-channel identity