A Wi-Fi handover, a tunnel, a lift. The connection dies mid-sentence and you are dropped onto an error screen, having lost the conversation with it. Every other kind of call rides that out. Now ours does too.

The trick is telling the two endings apart. A close the server actually decided — a budget stop, a sign-in problem — arrives with a reason attached. A transport death arrives with none. Only the second one retries: twice, quietly, with the interface held in "connecting" rather than flashing an error. The call resumes in the same conversation, with everything said so far still in it.

Waiting for you to finish

Whose turn it is used to be decided by silence alone. That is two defects in one: a quiet first syllable gets clipped off the front of a sentence, and a pause to think gets read as the end of your turn. It is now decided by what you are actually saying — the call waits when a sentence is obviously unfinished and closes promptly when it is not.

The per-minute charge moved with it. It is measured from your speech rather than from every second the microphone is open, so thinking costs nothing. The number on screen and the number on the bill come from the same measurement.

Loud enough to hear

People were holding the phone to their ear to hear a reply, which is not what a speakerphone is for. The spoken output now runs through a levelling stage that lifts quiet speech to a steady loudness without the crackle a flat volume boost produced. Separately, ending a call used to leave the whole app on the call output — read-aloud, music, video, all almost inaudible afterwards. The route is restored when the call ends.

And it reads what it writes

Voice could already save things to your long-term memory. It could not read them back, so it remembered facts it then did not know. A session now opens with what it has learned about you.