Victoria Song, writing at The Verge, tested the new Audio Intelligence features for Apple Watch that also rolled out yesterday as part of the iOS and watchOS 27.3 developer betas.
My first Siri Recap was for an editorial planning meeting for a story I’m working on. It was decent at identifying different topics discussed, and I was impressed that it stripped out some sensitive information. Any outsider looking at these notes would be terribly confused. Ironically, it meant that this summary wasn’t the most useful for my job.
What I will say is that I’ve been testing the Amazon Bee Frontier for the past few weeks. In the little time I’ve spent with the watchOS 27.3 beta, the differences between Audio Intelligence and the Bee are very clear. The latter has more of what I’d call an AI-forward presence. Bee, for example, offers insights about the emotional tone of a conversation. I’ve also found that the AI can be a bit judgy about insights and takeaways. Live Rewind is just a straight transcript. Apple’s said publicly that Siri Recap is meant to be highly factual notes to jog a person’s memory. That rings true based on my first one. And, as far as I can tell, there is no commerce angle for Siri Recaps, whereas Bee has an Amazon integration that surfaces buying suggestions based on conversations.
I forgot to link this, but a few weeks ago, Apple published a technical report (PDF link) that goes into more detail on how audio is recorded, transferred, and processed for Siri Recap:
The Secure Exclave on Apple Watch encrypts the audio and transmits it to the Secure Exclave on the paired iPhone. Once the transfer completes, audio is immediately deleted from the Secure Exclave on Apple Watch, and only the Secure Exclave of the paired iPhone can decrypt it.
And:
The encrypted audio is decrypted inside the Secure Exclave of the paired iPhone. On-device speech recognition transcribes the audio to text, and then an on-device language model generates a condensed version that is less than half the length of the original transcript. This is designed to distill the conversation to its core meaning, removing content suggestive of conversation tone, non-essential language, filler words, and redundancies, while preserving the topics discussed. Once processed, the raw audio is immediately and permanently deleted and no longer exists on either Secure Exclave. Only the condensed text remains. Finally, a safety model screens the text to omit potentially harmful terms.
The implementation of Siri Recap based on dedicated hardware and on-device processing sounds like quintessential Apple, and it seems like the feature does work as advertised. Alas, while I have access to Siri AI thanks to my American Apple Account and U.S. region setting, I live in Italy, and Siri AI does not speak Italian yet.
I do not recommend testing Siri Recap if your language isn’t officially supported by Siri AI – unless you’re looking for some laughs:
