Apple Begins Beta Testing Watch AI That Records 15 Seconds of Speech

Apple has shipped its first developer beta of "Audio Intelligence" for the Apple Watch Series 12, a suite of four opt-in AI features that can rewind and summarize human conversations. The code arrived in the iOS 27.3 update, roughly four months after Apple announced the technology at its Surprise and Shine hardware event in September.

The two features aimed at speech, Live Rewind and Siri Recaps, are the contested pair. They follow public backlash against AI recording wearables such as Meta's AI glasses. Apple's answer is engineering, not restraint.

How does Live Rewind work?

A double-press on the Watch's digital crown retrieves the last 15 seconds of conversation as a text snippet on the watch face. Apple frames the use case as missed information: verbal directions, a waiter reading specials, a fast-paced presentation.

The snippet is ephemeral by design. Unsaved snippets delete automatically after 30 seconds. Users can save them to the standalone Siri app that shipped with iOS 27. An "Ask Siri" button below the snippet answers follow-up questions, such as an unfamiliar technical term.

Apple also signals recording to bystanders. The double-press triggers an audible chime and an on-screen animation. Gesture, sound and display together tell the room that something was captured. Whether bystanders understand or notice this signal is another question Apple has not addressed.

What do Siri Recaps actually store?

Siri Recaps goes further than rewinding. It generates high-level notes of conversations a wearer participates in. Apple states the feature does not identify speakers, save raw audio, or produce word-for-word transcriptions. Instead, it compresses a conversation into a short sentence plus bullet points covering items such as dates, deadlines and pricing.

Apple suggests use in work meetings, parent-teacher conferences and training sessions. The recap, the company says, mimics a user's own note-taking without adding emotion or tone.

The retention rules are specific:

• Recaps delete themselves within seven days if the user does nothing.
• A daily review prompt appears, including on first morning unlock of the Watch.
• Users can save notes to the Notes app or export them to the Journal app.
• Recapping pauses during music playback or active noise cancellation.
• Users can run recaps manually or on a schedule, for example work hours only, and can disable them from Control Center at any time.

Apple says the system omits government-assigned identifiers, financial and authentication data, certain personal identifiers, and language promoting self-harm. That is a filtering claim. The beta testing now under way will show whether the filters hold in practice.

Which silicon is doing the listening?

The features run only on Apple Watch Series 12 and Apple Watch Ultra 4. They depend on the S11 chip, the built-in microphone, Apple Intelligence models running on the iPhone, and Apple's Private Cloud Compute.

The pipeline is layered. Lightweight AI models on the S11 chip detect whether a conversation involving the wearer is happening. Raw audio then flows into a secure enclave — a hardware-protected memory space the rest of the operating system cannot read. Only the paired iPhone can access that audio, through a secure pairing mechanism.

On the iPhone, on-device processing converts audio into condensed text with filler words removed. That text is encrypted and sent to Apple's Private Cloud Compute, which generates the recap and returns it. Apple says the recap and all intermediate summarizations are deleted from PCC immediately afterwards. Live Rewind uses the same watch-to-iPhone transfer path for its 15-second snippets.

The company claims the features will not affect watch or phone battery life. It offers no measured figures to support that claim. Ambient audio detection on an always-on processor carries a cost somewhere, and developers in this beta will be the first to measure it.

Two of the four Audio Intelligence features carry less controversy. Sound recognition alerts deaf and hard-of-hearing users to sirens, alarms and doorbells. Song recognition uses Shazam to name background music and its artist.

The developer beta precedes a public beta, then a general release. Apple has not committed to a date for either step, so the privacy guarantees and the battery claims remain unverified until real users test them outside Cupertino.