On-Device Dictation on the Mac: What Changed in macOS Tahoe

Where dictated audio used to go

Dictation looks local. You talk, words appear, and it feels like the Mac is doing the work. For most of the feature's history, it wasn't. When Apple shipped Dictation with OS X Mountain Lion in 2012, everything you spoke was recorded, sent to Apple's servers, transcribed there, and returned as text. No internet, no dictation.

That was the industry norm, not an Apple quirk. Speech models were too big for laptops, so vendors ran them in data centers, and dictated audio, some of the most intimate data a computer collects, became something you routinely shipped to a third party. Half-written emails, names, addresses, medical questions: all of it transited someone else's machines under whatever the retention policy said that quarter.

Many cloud dictation and transcription services still work this way, and the good ones document it clearly. But "we process your audio on our servers and we promise to be careful" is a policy. On-device is a property. A policy can change quietly; a property can be checked, with the Wi-Fi turned off.

Apple's slow move on-device

Apple started walking dictation back from the server early. OS X Mavericks (2013) added Enhanced Dictation, an optional local download that let the Mac transcribe offline. It was clunky, a hefty model per language and noticeably behind the server on accuracy, but it set the direction.

The real shift came with Apple silicon. The Neural Engine gave laptops the kind of compute that speech models used to need a rack for, and in recent macOS releases Apple made built-in keyboard dictation run on-device for supported languages on M-series Macs.

Developers stayed a step behind. The Speech framework's SFSpeechRecognizer (2016 on iOS, later on the Mac) defaulted to server-side recognition with duration caps, and its on-device mode was tuned for short utterances. Fine for "set a timer", rough for "take down this whole paragraph".

Years mark each approach's first appearance. On-device support in SFSpeechRecognizer and built-in dictation arrived gradually and varies by language and hardware.

What SpeechAnalyzer actually is

SpeechAnalyzer is the speech-to-text engine Apple introduced with macOS Tahoe (26). Strip the branding and three properties matter:

• It runs entirely on-device. The model lives on your SSD and executes on your Mac's silicon. Audio goes in, text comes out, and nothing crosses the network.
• It's built for long-form, real-world audio. Where the older developer engine was tuned for short commands, this one is designed for conversations, lectures, and extended dictation sessions.
• It's a public framework. Any Mac app can call it, the same way any app can use the system spell checker.

The mental model for a non-developer: your Mac now ships with transcription machinery inside the operating system. Apps ask macOS to listen and get text back, and the listening happens locally. Apple builds its own Tahoe transcription features on the same engine family, so this is core infrastructure, not a side project.

What this unlocks for third-party apps

Before Tahoe, a Mac app that wanted good dictation had two options: pipe your audio to a cloud API (quick to build, a bad answer to "where does my voice go"), or bundle its own local model, often a Whisper variant, and eat the download size, the memory, and the battery cost. Plenty of apps chose the cloud, and their users' audio went with them.

SpeechAnalyzer collapses that trade. The engine is already on the machine, maintained by Apple, updated with the OS. A third-party app can add private dictation with no model bundle, no cloud bill, and no privacy asterisk. When the honest option is also the cheap and easy one, more apps pick it. That is the quiet structural win of Tahoe's speech stack, and it matters more than any single feature built on it.

How NotchBay uses it (bias: I build it)

Bias on the table: NotchBay is my app, so this section is me showing my work, not reviewing myself.

NotchBay starts dictation from the notch: trigger it there, speak, and the text lands where you're typing. The notch is the one strip of screen that is always visible and never covered by a window, which makes it a natural handle for a system-wide tool. If you're new to the cutout, here's what actually lives inside the notch.

Under the hood it is SpeechAnalyzer end to end. The first build used SFSpeechRecognizer, and it showed: a short-command engine doing a long-form job. When Tahoe shipped, we replaced the whole path with the new engine, which is also why NotchBay requires macOS 26.

That matches the rest of the app's privacy stance: no account, no analytics in the app, everything local except actions you explicitly trigger. Among the notch apps compared in the 2026 roundup, dictation from the notch is a NotchBay angle rather than a category standard, so if this workflow is the draw, check for it specifically.

The honest limits

On-device dictation in 2026 is genuinely good, and it is not magic. The caveats worth knowing before you rearrange a workflow around it:

• macOS Tahoe only. SpeechAnalyzer does not exist on earlier releases. On older systems you're back to SFSpeechRecognizer or cloud services.
• Language support is Apple's call. The engine transcribes the languages Apple ships models for, and that list grows on Apple's schedule. I won't enumerate languages here because it would be stale by the time you read it: check your language in System Settings or in the app you plan to use.
• A one-time download. First use of a language may trigger a model download from Apple. After that, transcription is fully offline.
• Raw accuracy can still favor the cloud. A server-scale model with data-center compute can win on some accents and noisy audio. What on-device wins is the part no accuracy benchmark captures: your voice stays yours.

Frequently asked questions