The new generation of advanced AI dictation features Apple launched for iPhone 17 Pro and iPhone Air in the first iOS 27 developer beta are currently not turned on by default. Users need to manually turn it on in the system settings if they want to experience it.
Apple said that this new AI dictation system has a "significant improvement" in recognition accuracy. It is more reliable than the existing dictation system in real-time automatic processing of capitalization, punctuation, etc., and aims to significantly improve the overall voice input experience. This feature is driven by Apple’s latest AFM 3 Core Advanced model, a 20 billion-parameter, native multi-modal, sparse architecture system that activates only about 1 billion to 4 billion parameters at any one time based on specific requests to strike a balance between performance and computing power consumption.
To accommodate a model of this size on a smartphone device, Apple stores the complete model in flash memory instead of DRAM, and uses a lightweight routing module to select a fixed set of "expert" subnetworks during the initial processing stage and periodically reselect it during the generation process. This technology is called "Instruction-Following Pruning" by Apple. In Apple's internal parallel manual evaluation, AFM 3 Core Advanced was compared with the previous mass-produced dictation system in seven quality dimensions. The results showed that in the overall quality dimension, the new model gained a preference advantage at a ratio of 44.7% to 17.6%. This advantage was also basically consistent in the remaining six dimensions of punctuation, capitalization, format, semantic capture, spoken redundant processing, and style.
Due to the high model size and computing power requirements, the upgraded dictation function is currently limited to a few new devices, including iPhone 17 Pro, iPhone 17 Pro Max, iPhone Air, Apple Vision Pro equipped with M5 chips, iPads equipped with M4 and above chips and at least 12GB of running memory, and Macs equipped with M3 and above chips and at least 12GB of running memory. It’s worth noting that the standard iPhone 17 is excluded from the support list because it only comes with 8GB of running memory, which does not meet the 12GB threshold set by Apple for this feature. At the same time, Apple’s new generation of Siri voice with customizable expression styles is also based on this AFM Core Advanced model. It is also currently available as an optional preview function, which requires users to actively turn it on.
Apple emphasizes that the new dictation model runs entirely on the local device, so the transcription quality remains consistent whether the iPhone is connected to the network or not. This also means that the user's voice data does not need to be uploaded to the cloud, helping to enhance privacy protection. As for whether this advanced AI dictation feature will remain turned off by default when the official version of iOS 27 is released, or whether it will be automatically enabled in subsequent beta cycles, Apple has not yet given a clear explanation.