Advanced AI Dictation Not Enabled by Default in iOS 27 Beta

[MacRumors]

Apple's next-generation AI dictation feature for the iPhone 17 Pro and iPhone Air is not turned on by default in the first developer beta of iOS 27.

Apple says the new AI-powered dictation system delivers "a major boost in accuracy," with more reliable on-the-fly capitalization and punctuation than the existing dictation system. The feature runs on Apple's new AFM 3 Core Advanced model, which is a 20-billion-parameter, natively multimodal system that uses a sparse architecture, activating just one to four billion parameters at a time depending on the request.

To fit a model that large onto a smartphone, the full model is stored in flash memory rather than DRAM, with a lightweight routing block selecting a fixed set of "experts" during initial processing and periodically reselecting them during generation, a technique Apple calls Instruction-Following Pruning.

In side-by-side human evaluations against Apple's previous production dictation system across seven quality dimensions, AFM 3 Core Advanced was preferred on overall quality by a margin of 44.7% to 17.6%, with that preference holding consistently across the other six dimensions, which include punctuation, casing, layout, meaning capture, disfluency handling, and style.

Because of the model's size, the upgraded dictation is limited to a handful of newer devices: the iPhone 17 Pro and iPhone 17 Pro Max, the iPhone Air, the Vision Pro with M5 chip, iPads with an M4 chip or later with at least 12GB of RAM, and Macs with an M3 chip or later with at least 12GB of RAM. Notably, the standard iPhone 17 is excluded, as it ships with 8GB of RAM rather than the 12GB the larger model requires. The same AFM Core Advanced model also powers Apple's new customizable expressive Siri voices, another opt-in preview as of beta 1.

The new dictation model runs entirely on-device, so transcription quality stays the same whether or not the iPhone is connected to a network. It remains unclear whether the preview will stay off by default when iOS 27 is released officially later this year, or whether Apple will switch it on automatically at some point during the beta cycle this summer.

Article Link: Advanced AI Dictation Not Enabled by Default in iOS 27 Beta

 

[Lounge vibes 05]

It’s also not available within macOS just yet, even for compatible machines.

 

[mrhansolo]

It’s also not available within macOS just yet, even for compatible machines.

Thank you for adding that comment. After reading the article that's what I was going to ask. I used my M4 macbook pro for a journal entry yesterday and did not notice any difference in dictation. Glad to know why now.

 

[Astuces iOS]

Well it will come today in beta 2 !

 

[CDPlayer]

I just hope Apple allows users to turn it all of as they can with 26. I’m sure future beta’s will add that ability… right? … right?!

 

[Elwe]

I can’t imagine that the next generation of base iPad will have 12GB of RAM. This is another reason, combined with the much higher price of the iPad Pro, for the iPad Air models to continue to exist.

On-device or via a Cloud, it seems to me AI accuracy is going to become even more meaningful a differentiator. So a key question for the devices which don’t have this amount of memory is how good will the free or very cheap Cloud models they can tap into be? Apple hasn’t yet been forthcoming, for example, of what exact restrictions their Cloud tiers will impose. As OpenAI, Anthropic, Google, etc, all dramatically increase the price of their offerings to cover the actual CoGS, “ a good-enough default” offering by Apple at reasonable cost could take real market share.

 

[thelion7]

Apple's next-generation AI dictation feature for the iPhone 17 Pro and iPhone Air is not turned on by default in the first developer beta of iOS 27.

Apple says the new AI-powered dictation system delivers "a major boost in accuracy," with more reliable on-the-fly capitalization and punctuation than the existing dictation system. The feature runs on Apple's new AFM 3 Core Advanced model, which is a 20-billion-parameter, natively multimodal system that uses a sparse architecture, activating just one to four billion parameters at a time depending on the request.

To fit a model that large onto a smartphone, the full model is stored in flash memory rather than DRAM, with a lightweight routing block selecting a fixed set of "experts" during initial processing and periodically reselecting them during generation, a technique Apple calls Instruction-Following Pruning.

In side-by-side human evaluations against Apple's previous production dictation system across seven quality dimensions, AFM 3 Core Advanced was preferred on overall quality by a margin of 44.7% to 17.6%, with that preference holding consistently across the other six dimensions, which include punctuation, casing, layout, meaning capture, disfluency handling, and style.

Because of the model's size, the upgraded dictation is limited to a handful of newer devices: the iPhone 17 Pro and iPhone 17 Pro Max, the iPhone Air, the Vision Pro with M5 chip, iPads with an M4 chip or later with at least 12GB of RAM, and Macs with an M3 chip or later with at least 12GB of RAM. Notably, the standard iPhone 17 is excluded, as it ships with 8GB of RAM rather than the 12GB the larger model requires. The same AFM Core Advanced model also powers Apple's new customizable expressive Siri voices, another opt-in preview as of beta 1.

The new dictation model runs entirely on-device, so transcription quality stays the same whether or not the iPhone is connected to a network. It remains unclear whether the preview will stay off by default when iOS 27 is released officially later this year, or whether Apple will switch it on automatically at some point during the beta cycle this summer.

Article Link: Advanced AI Dictation Not Enabled by Default in iOS 27 Beta

Tim saving those tokens to buy more RAM…

 

[Obioban]

Tim saving those tokens to buy more RAM…

runs on device, so costs apple nothing.

 

[Skyuser]

Tim saving those tokens to buy more RAM…

These models do not consume any token from Apple's servers because it runs entirely on-device.

 

[david0128]

This seems like a case of planned obsolescence on the Mac front. I have an M2 MacBook Pro Max with 96gb of ram - it seems quite odd that it's not powerful enough to run the same AI models as a base M3, unless Apple just wanted to limit the amount of machines they had to test it on.

 

[svish]

Good to know about this. Anyway will try this out on my 17 Pro Max. RAM requirement is not a surprise for Apple Intelligence.

 

[Ahheck01]

Wait, not turned on (meaning off by default), or not yet available? I haven't been able to find the option on my 17 Pro Max.

 

[GfPQqmcRKUvP]

I opened the link and I don’t think people should be excited or jealous about this new dictation model.

In more than 50% of examples a human reviewer found the previous dictation model/capability as good or better than the new dictation model. There are about to be a LOT of underwhelmed people.

 

[ry-guy]

side-by-side human evaluations against Apple's previous production dictation system across seven quality dimensions, AFM 3 Core Advanced was preferred on overall quality by a margin of 44.7% to 17.6%, with that preference holding consistently across the other six dimensions, which include punctuation, casing, layout, meaning capture, disfluency handling, and style.

I’ll wait to see how all the new Pro phone users fare if they think it really makes a difference. I doubt it’s worth the price of the new phones.

 

[CDPlayer]

This seems like a case of planned obsolescence on the Mac front. I have an M2 MacBook Pro Max with 96gb of ram - it seems quite odd that it's not powerful enough to run the same AI models as a base M3, unless Apple just wanted to limit the amount of machines they had to test it on.

Even the M1 Ultra should have no problems with it. I don’t want it but it’s absolutely a move to push more Mac sales which is even worse as AI is the reason for the shortage and price increases.

 

[Lounge vibes 05]

Wait, not turned on (meaning off by default), or not yet available? I haven't been able to find the option on my 17 Pro Max.

It’s in settings, under general, and then under keyboards.

 

[gnomeisland]

Even the M1 Ultra should have no problems with it. I don’t want it but it’s absolutely a move to push more Mac sales which is even worse as AI is the reason for the shortage and price increases.

I think it is a question of *how* the foundational models are run. Likely they are coded to run exclusively on the Neural Engine. There were pretty beginning leaps between M2 & M3 (and again between M4 & M5 but it seems not enough to warrant a separate tier in this case).

I agree, it would be nice to have that option for those of us running higher end M1 & M2 Macs.

 

[rageguy]

Isn't this the setting where you go to Settings, General, Keyboard, and enable Advanced Dictation Preview? For eligible iPhones on iOS 27 DB1.

 

[ifxf]

runs on device, so costs apple nothing.

Generating the model and keeping it up to date is a cost center.

 

[Nick_P]

ChatGPT gets my voice perfectly, even when I stutter, on my “old” iPhone 14. It completely blows Siri out of the water. Idk why Apple didn’t do the same thing and instead is doing everything on-device hogging all of the local resources.

If the new voice recognition is only preferred 44% to the old one’s 17%, that means it must still be pretty bad.

I’d imagine if the same test was done with ChatGPT and current Siri, it would be 99.999% to .001%.

 

[CDPlayer]

I think it is a question of *how* the foundational models are run. Likely they are coded to run exclusively on the Neural Engine. There were pretty beginning leaps between M2 & M3 (and again between M4 & M5 but it seems not enough to warrant a separate tier in this case).

I agree, it would be nice to have that option for those of us running higher end M1 & M2 Macs.

I just can’t believe an M1 Ultra can’t run it but an iPhone 17 Pro and Air can. I know, slight neural engine differences. I’m the last person to state such but if Apple wanted to support M1 and M2 systems they could. As macOS 27 is beginning to drop some support, I suspect Apple is preparing to sunset older M systems which hurts as they skipped the M4 Ultra and most likely will with the M5 otherwise I would have updated already. Ironically, AI will make the next iteration eye watering expensive compared to now.

I hate this timeline.

 

[John Connor]

Call me old-fashioned, but you would expect the article to advise how to enable it. Rather than all the waffle.

 

[CapitalIdea]

Tim saving those tokens to buy more RAM…

Read the article, and if parts of it are unclear it is just as easy to ask as it is to proudly declare something that is fundamentally nonsensical.

 

[Siliconguy]

This seems like a case of planned obsolescence on the Mac front. I have an M2 MacBook Pro Max with 96gb of ram - it seems quite odd that it's not powerful enough to run the same AI models as a base M3, unless Apple just wanted to limit the amount of machines they had to test it on.

The difference between M2 and M3 NPU performance is not that large. So it is a bit odd to have the cutoff there. M4 was a big jump in NPU performance.

 

[tedpearlman]

It’s in settings, under general, and then under keyboards.

Thanks. I was wondering. 😀