Apple Watch's conversation recaps hit beta, with a chime to tell people you're rewinding them

Apple's "Audio Intelligence" features for the Apple Watch are now in beta for developers in iOS 27.3, two months after Apple announced them at its September hardware event, as TechCrunch's Sarah Perez reported. Two of the four features are the interesting ones: they listen to the people around you.

The set works on the Apple Watch Series 12 and Ultra 4, and everything is opt-in. Sound recognition flags sirens, alarms and doorbells for deaf and hard-of-hearing users, and song recognition uses Shazam. The other two are Live Rewind and Siri Recaps.

What actually shipped

Live Rewind is a 15-second undo button for your ears. Double-press the digital crown and the watch shows what was just said as text: the waiter's specials, a list of directions, a name you missed. An "Ask Siri" option lets you ask about what you caught. The snippet disappears after 30 seconds unless you save it to the Siri app.

Siri Recaps goes further: it summarizes whole conversations into a sentence and a few bullet points, such as dates, deadlines and prices. Apple says it doesn't identify speakers, keep raw audio or produce word-for-word transcripts. You can run it manually or on a schedule, such as work hours only. It pauses while you're listening to music on headphones. At the end of the day the watch asks which recaps to keep, saved to Notes or Journal. Anything you ignore deletes itself within seven days. Apple says recaps leave out government identifiers, financial and login details, some personal identifiers and self-harm language.

The architecture is the interesting part

The pipeline is split across three places. Small models on the watch's S11 chip detect that a conversation involving the wearer is happening. The audio goes to a secure enclave on the paired iPhone, which turns it on-device into "condensed text" with the filler words stripped. Only that text, encrypted, goes to Apple's Private Cloud Compute to be summarized. Apple says the raw audio isn't accessible to anyone, Apple included.

It's a careful design, and a useful template for anyone building ambient features: detect on the cheapest chip, reduce on the device, and send the cloud as little as possible.

The people on the other side

The consent model is thinner than the cryptography. When you rewind, the watch plays a chime and shows an animation so people nearby know they were captured. That only works if they hear the chime and know what it means. Apple is clearly trying to avoid the backlash Meta's AI glasses and other AI wearables got. Whether a chime is enough will be settled by the first parent-teacher conference where somebody notices.

Scoreboard update: AI +1. Your watch can now take the meeting notes, whether or not the meeting knows.

Stay undead.