How Apple Dictation Works: On-Device, Server and AI

Apple Dictation converts speech to text with a neural speech model, and where that model runs depends on your Mac and the language you dictate in. On an Apple silicon Mac, Apple lists 48 of its 63 Dictation language and regional variants as on-device, so your words are converted on the Mac and not sent to Siri's servers ^{1 2 3}. Apple's macOS Sonoma 14 guide says that Intel Macs, and languages without on-device support, send your words to Apple ⁴. The Keyboard settings pane shows which route you are on.

Accuracy is a separate question. The one independent test we found put Apple's speech engine at a 10.3 percent word error rate on a single recording in 2025 ⁵, and that test measured a developer interface, not the Dictation key itself. The sections below cover each part, from the overview of dictation on Mac to the privacy guide on what Apple keeps.

Apple Dictation turns speech into text with neural models, on the Mac or on Apple's servers

Yes, Dictation uses AI in the sense that matters for speech recognition: a machine-learning model turns the sound of your voice into written words. We did not find a published technical paper on the Dictation model itself. Apple's 2017 article on the "Hey Siri" detector does describe the method, and it says the detector "uses a Deep Neural Network (DNN) to convert the acoustic pattern of your voice at each instant into a probability distribution over speech sounds" ⁶. That is the first stage of a speech model: sound goes in and probabilities over speech sounds come out.

A recognizer then chooses the most likely words from those probabilities. An acoustic model scores the sounds, and a language model favors word sequences that make sense in your language. Punctuation and capital letters are added on top. The simplified path looks like this:

Apple's newest addition is an Apple Intelligence model. Apple says its improved on-device Dictation model needs a Mac with M3 or later and at least 12GB of unified memory ⁷.

So the direct answer is yes: the core is a neural speech model, and macOS 27 adds a newer Apple Intelligence model for some English users. Whether a feature is called AI matters less than where it runs, which is the next question. We answer it in more detail on the question of whether Dictation uses AI.

Keyboard settings say whether your dictation stays on the Mac

Apple's Mac Help says you can check whether your voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers, and that the check lives in Keyboard settings ³.

To check on your own Mac:

1. Choose Apple menu > System Settings, then click Keyboard in the sidebar, as described in Apple's Dictation guide.
2. Find the Dictation section and read where it says your dictation is processed. Apple says the Mac shows that status there.
3. If the status says the words go to Apple, you are on the server route. Apple's Sonoma 14 guide says a language without on-device support goes to Apple, so check the language you dictate in as well.

The two routes differ like this:

Storage is a separate setting from routing. The Share Audio Recordings choice decides whether Apple may store audio of your Siri and Dictation interactions, and Apple may review a sample of what it stores ⁹. Choosing Not Now declines. The privacy guide goes through each choice.

On-device dictation works offline; server dictation needs a connection

Yes, for languages on the on-device list, although Apple's wording has changed over the years. Apple's macOS Monterey 12 guide said that on Apple silicon, general text dictation was processed on the device in many languages, and that no internet connection was required ¹⁰. Later guides are less direct. The Dictation troubleshooting page says you may need an internet connection, and tells you to check Keyboard settings to find out ¹¹.

Server dictation cannot work offline because the speech model is not on the Mac. Apple's 2017 article describes the same split for Siri: the audio reaches the server only if the stages on the device pass it on ¹². The practical rule, as Apple's Monterey guide puts it, is this: if the language is on the on-device list and your Mac has Apple silicon, dictation keeps working when the network drops. Otherwise, it will not. The route also changes how dictation feels. On the on-device route, nothing waits on a network round trip. On the server route, every phrase depends on your connection, so a weak signal shows up as lag between speaking and text appearing. Our guide to offline dictation on Mac covers the offline options in more detail, and the question on offline use gives the short answer.

Apple silicon and macOS version decide which model you get

Two things set the route: the chip and the language. The table below shows what Apple's Mac guides said in four releases. The wording changed from one release to the next, so the Keyboard settings on your own Mac are the authority.

Apple's machine learning team gives privacy as a reason for on-device processing in general. It says that on-device machine learning benefits privacy, because data for inference workloads remains on the device rather than on a server ¹⁴. Apple's Mac Help does not name the processor that runs Dictation, so this article does not claim that the Neural Engine runs it.

Hardware sets the upper limit. Tahoe 26 is the last macOS release that still lists Intel Macs in our sources, so an Intel Mac cannot get the macOS 27 dictation model ¹⁵.

Language coverage is the other limit. Apple lists 63 Dictation variants in total ², and 48 are on the on-device and modeless list ¹. The 15 that are not on that list are the ones to check in Keyboard settings before you rely on offline use ¹⁶.

Apple warns that Dictation is not available in all languages or regions, and that features may vary ¹⁷. Check the full list of Dictation languages on Mac before you choose a language for a long document.

Accuracy depends on your microphone, your language setting and how you speak

The one independent test we found is a 2025 review by 9to5Mac. Its author transcribed one 7-minute-31-second recording with Apple's speech API, and got a word error rate of 10.3 percent ⁵. The same recording gave 1.5 percent with Whisper Large V3 Turbo, and 12.3 percent with NVIDIA's Parakeet v2 ^{18 19}.

Read that result with three limits in mind. First, the author reached Apple's speech API through an open-source command-line tool on an M2 Pro MacBook Pro with 16GB of RAM ²⁰. The test therefore measures the developer speech interface, not the Dictation key, and the sources we read do not say the two use identical models. Second, word error rate also depends on how you normalize casing and punctuation, so a different method gives a different figure. Third, it is one recording. Treat 10.3 percent ⁵ as a single measurement, not a rating of every Dictation session. We found no other independent test of Apple Dictation in 2026. Vendor blogs publish figures for it, but they come from companies that sell rival apps, and we could not check their methods, so we leave them out. The question on whether Apple Dictation is any good puts this evidence in context.

You can run a rough test yourself in about ten minutes. Read a 200-word passage of your own writing into a blank document with Dictation on. Count the words that came out wrong, including punctuation you wanted and did not get, and divide by 200 for a rough error rate. Repeat it with the microphone you normally use and with the built-in one, to see how much the microphone matters for you.

Your microphone matters as much as the model. Apple advises you to position yourself as you normally use your Mac, speak clearly at a volume that is neither too quiet nor too loud, and avoid background noise ²¹. Distance is the largest single effect in the research we checked. In the Ravanelli et al. 2017 figures, a single distant microphone gave a word error rate about 3.2 times the close-talking rate ²². The two conditions are not identical, so treat that ratio as rough.

Your language setting decides which model handles your speech. You can change the dictation language, or set up several languages and switch between them ²³. Apple describes switching between languages, not recognizing two languages in one sentence. If you dictate a mixed-language sentence, do not expect it to come out right, and check the output. The troubleshooting guide covers the language and microphone checks in order.

Apple Dictation has no voice training, but Voice Control lets you add your own words

No, Apple Dictation has no voice-training step. In the Dictation pages we read, Apple describes no exercise where you read text aloud so the software learns your voice, and it describes no custom word list for Dictation. Dragon for Mac is different. In Nuance's help for Dragon for Mac version 6, Voice Training asks you to read aloud a story so the software learns how you speak ²⁴. That is the step many people expect, and it is missing from Apple's Dictation.

Apple does not store dictation sent to its servers unless you opt in ⁸. If you choose Improve Siri and Dictation, Apple can store and review a sample of your Siri and Dictation audio ²⁵. Apple's pages do not describe a personal voice profile built from that audio. The question on whether Dictation learns your voice gives the short answer, and the glossary entry for voice profile explains the term.

If you need custom words, use Voice Control instead. Voice Control accepts up to 1,000 vocabulary terms, and you can teach it how to say each one ²⁶. To add them, choose Apple menu > System Settings, click Accessibility in the sidebar, then click Voice Control, and follow the steps for adding individual vocabulary terms in Apple's custom vocabulary guide. The trade-off is that while Voice Control is on, you dictate through it, and standard macOS Dictation is not available ²⁷. The step-by-step question on adding words covers the setup, and the guide to Voice Control on Mac covers the commands. Add the words that a general model is least likely to know first: names, product names and technical terms, each with the pronunciation you actually use.

Third-party apps can use the same engine through Apple's Speech framework

Apple's Speech framework lets third-party apps add speech-to-text using Apple's own models ²⁸. The newer SpeechAnalyzer class, available from macOS 26, transcribes entirely on the device, but the language models must be downloaded first ²⁹. The older SFSpeechRecognizer could use Apple's servers on devices with limited resources ³⁰.

This is one route a third-party app can take to reach Apple's speech models. The accuracy evidence above applies to them only loosely, because the 9to5Mac test measured the developer interface, not any particular app. An app built on the framework is judged by its own settings, its own text handling and its own privacy policy. Yap, which we make, is an on-device dictation app for Mac and Windows, and its download page describes how it is installed. When you evaluate any dictation app that claims on-device processing, read its privacy policy for an explicit statement of where audio goes, and test it with the network turned off. That test shows the route more reliably than marketing copy. For most people, the built-in tool is enough: it comes with macOS and works for messages and notes, but Apple says dictating in a search box is not part of general text Dictation ³¹. The comparison of dictation and Voice Control and Siri on Mac helps you choose between them.