Apple Details Siri's Privacy through Secure Exclave

Apple explains how its new audio AI features process raw audio locally within a Secure Exclave of the S11 chip in the Apple Watch Series 12 and Ultra 4, without recording or external access. Activation depends on the user, and the announced transfers or synchronizations are encrypted under specific conditions.

On-Demand Activation and Conditional Synchronization

Apple specifies that the user retains control over the activation of Audio Intelligence features. For instance, Live Rewind requires double-tapping the Digital Crown each time it is used. If Audio Intelligence needs to be transferred to the iPhone, this transfer occurs encrypted via the Secure Exclaves of both devices. Users also choose which texts from Siri Recap and Live Rewind they wish to keep. The synchronization of these texts benefits from end-to-end encryption, provided that the device has a passcode and that iCloud two-factor authentication is enabled.

Local Microphone Processing Without Access by watchOS, Apps, or Apple

The raw audio captured by the microphone is processed in dedicated hardware, the Secure Exclave, without being recorded as a file. This initial processing allows for the analysis of speech, sounds, or music, but does not result in any transcription or retention of the raw audio. The audio buffer functions as a continuously overwritten stream, which remains confined within the protected hardware and produces no recordings. The data processed in this environment is not accessible by watchOS, apps, the user, or Apple.

Announced Features and Support for the S11 Chip in the New Apple Watches

At the launch of the iPhone Duo, Apple introduced several audio AI features for Siri: Siri Recap, Live Rewind, Sound Recognition, and Music Recognition. These features rely on the Secure Exclave integrated into the S11 chip, which powers the Apple Watch Series 12 and Apple Watch Ultra 4. This isolated hardware compartment within the silicon allows for separate processing of sensor data. Apple claims it cannot access the raw audio recordings from these features and has published a document detailing its approach to balancing ambient listening with privacy respect.