Apple is testing a new AI-powered dictation system in iOS 27 that dramatically improves transcription accuracy, according to details discovered in the first developer beta.
Dictation has long been one of those features that work well enough until they don’t. A quick text message might come through perfectly, while a longer note can end up filled with misplaced punctuation, odd capitalization, or words that sound similar but carry an entirely different meaning. Apple’s new model appears to be focused on reducing those kinds of mistakes.
In Apple’s internal evaluations, human reviewers preferred the new system over the current dictation experience by a wide margin across multiple quality categories.
The model is better at handling natural speech patterns, including pauses, interruptions, and sentences that aren’t perfectly structured before they’re spoken aloud.
The technology behind it is notable in its own right. Apple is reportedly using a 20-billion-parameter model that runs directly on the device rather than in the cloud.
Instead of loading the entire model into memory at once, the system pulls in only the portions it needs at a given moment, allowing it to operate on consumer hardware.
That approach comes with hardware requirements. The feature currently requires at least 12GB of RAM, limiting support to a relatively small group of devices.
The list includes iPhone 17 Pro, iPhone 17 Pro Max, iPhone Air, Vision Pro with M5, supported M4 iPads, and Macs running M3 or later with sufficient memory.
Because the model runs locally, dictated text never leaves the device for processing. That means it continues to work without an internet connection and keeps voice data on the user’s hardware rather than sending it to remote servers.
The same model is also being used for Apple’s new customizable Siri voices, which appear as an optional preview in the current beta.
Apple has not said whether either feature will be enabled by default when iOS 27 launches later this year.
