Apple held WWDC 2026 on June 8 and 9, and the headline was AI: a rebuilt Siri, a new generation of on-device Foundation Models, and — said out loud on stage — "higher-accuracy dictation." If you dictate on your Mac, that last part is the line worth paying attention to.

So here is the honest question this post answers: did Apple just make a dedicated dictation app pointless? Short version — no. It raised the floor. The built-in baseline got better, which is good for everyone, but the things that send people looking for a dedicated tool in the first place mostly weren't on stage. Here's what changed and what didn't.

What Apple actually announced

A few things are real and confirmed, separate from the marketing gloss.

Siri AI. Apple introduced "an entirely new version of Siri deeply integrated into iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro." It's conversational, has its own standalone app that syncs your history over iCloud, can answer questions about what's on your screen, pull context from your messages, emails, and photos, and take actions across apps. It ships as a beta later in 2026, English first. There are real launch caveats: in the EU it arrives on Mac and Vision Pro but, in Apple's own wording, "not initially in the EU in iOS, iPadOS, and watchOS," and it isn't coming to China at launch while Apple works through regulatory requirements.

Third-generation on-device models. The dictation improvement comes from here. Apple's on-device lineup is now AFM 3 Core, a 3-billion-parameter dense model, and AFM 3 Core Advanced, a 20-billion-parameter sparse model that activates only 1 to 4 billion parameters per request and is natively multimodal. Apple credits that Advanced model specifically with "expressive voices and higher-accuracy dictation," and reports human raters preferred its overall quality 44.7% to 17.6% over the previous system. That's a genuine step up, running on the Neural Engine.

The Gemini footnote. This one gets misreported, so it's worth being precise. Apple and Google announced a multi-year deal under which "the next generation of Apple Foundation Models will be based on Google's Gemini models and cloud technology." But Apple was equally clear that the models shipping on your device contain "none of the models that Google deploys" — Gemini was used to help train and distill Apple's models, not to run on your Mac. Worth knowing, because the privacy story below depends on it.

For developers there's more: Apple opened its Foundation Models framework behind a new Swift LanguageModel protocol so apps can swap between Apple's on-device model, cloud Gemini, Anthropic's Claude, or community MLX models with a one-line change, and shipped Core AI, a local inference framework that runs across CPU, GPU, and Neural Engine "with no server and no cost per token." That direction matters more than any single feature, and we'll come back to it.

The genuinely good news

Give Apple the credit it's due. On-device dictation accuracy improving, for free, private by default, with zero setup, is a real win. If you dictate the occasional message or note into a text field and the only thing that ever bothered you was the odd misheard word, macOS just got better at exactly that, and you may not need anything else. That's the honest baseline.

The reassurance most of this post is about isn't "Apple's update is weak." It's better than last year's. It's that "better dictation accuracy" and "a smarter assistant" are not the same job as the workflow a dedicated app is built for.

Where it still doesn't reach

Here's what wasn't on the WWDC 2026 stage, framed honestly as what Apple did and didn't announce.

An assistant is not a dictation tool. Siri AI is the big swing, and it's an assistant: ask it things, have it take actions, hold a back-and-forth. That's a different job from voice typing — getting your exact words into the exact app and field your cursor is in, whether that's your editor, Slack, a code comment, or a support ticket. Apple made the assistant much better. It didn't show a system-wide voice-typing layer that drops clean text wherever you're working.

Meetings and speakers. Nothing at WWDC 2026 captured a Zoom or Google Meet call's system audio and split the transcript by who was talking. Apple did not announce on-device speaker diarization. If you transcribe meetings and need "Alice said / Bob said" labels without a bot joining the call, that's still a job for a dedicated tool. We wrote about transcribing meetings on Mac without the cloud separately.

Translation while you dictate. Improved dictation is about getting your speech into text accurately. Speaking French and getting clean English at your cursor, in whatever app you're in, is a separate pipeline Apple didn't put on stage. More on how that local translation pipeline works if you need it.

Engine choice and files. Dedicated local apps let you pick your speech engine — Whisper Large-v3 for accuracy, Parakeet for speed — and transcribe existing audio files, not just live speech. Apple gives you Apple's model. For most people that's fine. For the people who care, it's not a choice they get. See our Whisper vs Parakeet comparison for why the engine matters.

The privacy nuance worth reading twice

Apple's on-device model is genuinely private — it runs on your Mac and the audio doesn't leave. No argument there. The nuance is the tiers above it. Heavier requests go to Private Cloud Compute, which Apple this year extended onto NVIDIA Blackwell GPUs running in Google Cloud, and the next-generation models are trained with Gemini. Apple says your data isn't stored or made accessible to Apple or anyone else, and that Google never sees it. Those are Apple and Google's own assertions about their own systems, and reasonable people can decide how much that's worth.

If your bar is "everything stays on this machine, no cloud tier, no trust required," a fully-local tool still clears a line Apple's tiered architecture, by design, does not. That's the whole reason offline, on-device voice to text exists as a category, and WWDC 2026 didn't change the math on it.

So do you still need a dedicated app?

Honest answer, both directions:

  • Probably not, if you dictate occasionally into text fields and want something free and built-in. macOS 27's improved on-device dictation is a real upgrade and it's right there. Use it.
  • Still yes, if you voice-type all day across every app, transcribe meetings with speaker labels, translate as you speak, want to choose your engine, or need a hard guarantee that nothing leaves your Mac. Those are the jobs Apple didn't ship.

For that second group, Vext is one option built for exactly it: system-wide dictation into any app, meeting transcription with speaker labels, live translation, and voice notes, all running on local Whisper or Parakeet plus a local LLM for cleanup, $49 once, no subscription. The honest tradeoffs: it's not free, it's Apple Silicon only, and now that Apple's baseline is better, casual users genuinely may not need it.

The bigger signal

The most interesting thing at WWDC 2026 wasn't any one feature. It was Apple shipping Core AI and opening on-device models to every app, betting that the right place to run AI is on the silicon you already own. That's the exact thesis dedicated local voice apps were built on. Apple didn't end that category this year. It validated it — and raised the floor underneath it.