How can I improve my dictation accuracy?

Fix the audio first, because it is the cheapest and largest gain. Get a microphone within a hand-span of your mouth, which raises the level of your direct voice against room reflections and background noise. Dictate in a room with soft surfaces rather than bare walls and glass. Speak in complete phrases at your normal pace instead of one exaggerated word at a time, because the model uses the surrounding words to decide what it heard. Then set your dictation language to match the way you actually speak. Only after all of that does changing the speech model itself become the thing that matters.

How do I improve my Mac microphone for dictation?

Move it closer and point it at your mouth. The built-in microphone sits in the laptop body, so it hears your voice after it has crossed the room and bounced off the walls, which is the condition speech researchers call far-field and treat as a separate, harder problem than close-talk audio. Any wired headset, a pair of earbuds or a desk microphone within a hand-span of your mouth removes most of that. Then check that the right input is selected in System Settings under Sound, and that the input level moves clearly when you talk. One thing that will not help: the Voice Isolation mic mode is set inside an app that supports it, such as FaceTime, so it is not a switch you can apply to Dictation.

Can I add custom words to Mac dictation?

Not to Dictation itself. Apple does provide a custom vocabulary of up to 1,000 terms, for things like CAD commands or medical words, but it is a Voice Control feature, so you reach it in System Settings under Accessibility and then Voice Control. The catch is that when Voice Control is on you cannot use Dictation, so you are choosing between the two rather than adding vocabulary to the dictation you already use. If proper nouns and technical terms are the bulk of your errors, changing the speech model is a more direct fix than working around the vocabulary gap.

What is the most accurate dictation software?

There is no single winner, and the honest reason is that accuracy depends more on your setup than on the software. Three things decide it: how close the microphone is to your mouth, whether the model has met the names and jargon you use, and how well it matches your accent. A modern local model such as Parakeet or a recent Whisper build will usually beat Apple Dictation on punctuation and proper nouns, but Apple publishes no accuracy figure for Dictation, so there is nothing official to compare it against. Published benchmark scores come from clean read-aloud audio, not from a laptop in a noisy room.

← Resources

How to improve Mac dictation accuracy

Short answer: the levers that move Mac dictation accuracy, in order of payoff, start with getting a microphone close to your mouth, then cutting room reflections and background noise, then setting the dictation language to match how you speak, then speaking in complete phrases at your normal pace. Those four get you most of the way, and there are five more below them. What none of them fix is the model, because Apple Dictation has no custom vocabulary and does not learn your voice. This page is about quality: dictation works, but the output is not good enough. If it is broken instead, start with dictation not working on Mac, or with Apple Dictation cutting off mid-sentence if it stops on its own. The wider category map is the guide to local speech-to-text on Mac.

Why is Mac dictation so bad?

Mac dictation feels bad because you cannot teach it anything. Apple's built-in speech to text has no custom vocabulary, no way to train it on your voice, and no published accuracy figure to hold it to. It has to serve dozens of languages and every text field in macOS, so it is tuned for the average speaker saying average things. Your surnames, drug names, case references and product names are not average.

Four specific gaps explain most of the frustration, and it is worth separating them because only some of them have fixes.

You cannot add a word. Apple does ship a custom vocabulary of up to 1,000 terms, with examples like the CAD command DIMSTYLE or the medical term hemianopsia. It belongs to Voice Control, not to Dictation. There is no equivalent panel for Dictation, so a proper noun is whatever the general model guessed.
It does not learn your voice. The Improve Siri and Dictation switch reads like a personalization setting. It is not. Apple describes it as an opt-in where you agree to let it store and review audio and transcripts to generally improve Siri, Dictation and Search. You are donating samples to the next model, not training yours.
Where it runs varies. Apple tells you to check Keyboard settings to see whether your configuration needs an internet connection or processes your voice on the device. That answer changes with your language and your Mac, so two people can get different behavior from the same feature. The detail is in whether Apple Dictation is on-device.
There is no number. Apple publishes no word error rate and no model size for Dictation. Any accuracy percentage you see quoted for it is somebody's own test on their own voice, not a spec. If you want to know how to read an accuracy claim at all, that theory sits in word error rate explained; this page stays on the practical side.

None of that makes the feature useless. It makes it a general-purpose tool with no adjustment surface, which is exactly why the levers that do exist are all in the audio and the settings around it.

Three things the older guides still get wrong

Most of the advice ranking for this topic was written years ago and never revisited. Three claims in particular are now stale, and believing them sends you chasing fixes that do not exist.

Enhanced Dictation was not simply renamed. MacPaw's guide says Enhanced Dictation "was renamed in macOS Catalina" and still works as it did in older versions. It was removed. Howard Oakley documented the change as it happened: the Dictation tab stayed in the Keyboard pane but the enhanced offline engine went. Apple's own discussion threads are full of Monterey users hunting for the checkbox and being told it is gone. So there is nothing to re-enable.
The old length cap is gone. Plenty of pages still repeat a hard limit of under a minute per pass. Apple's current documentation says you can dictate text of any length without a timeout, and that dictation stops automatically after 30 seconds of no detected speech. Those are two different behaviors, and only the second one is still live. The silence stop has its own piece: why Apple Dictation keeps cutting off.
The formatting commands came back. The Catalina-era complaint was losing spoken formatting. Apple's current command reference documents new line, new paragraph, caps on and off, all caps on and off, numeral, roman numeral, no space on and off, and tab key. The full set is in Mac dictation commands.

The accuracy levers, and what each one is worth

Here is every lever, ranked by how much it typically buys you, with the point at which each one stops helping. Work down the list rather than jumping to the bottom row, because the top three are free and the bottom one is not.

Lever	Why it works	Roughly what it is worth	Where it runs out
Microphone distance	Close-talk audio carries far more of your direct voice and far less reflected sound	The largest single gain, and free if you own headphones	A clean signal into a general model still misreads names and jargon
The room	Hard walls and glass send late reflections back into the mic, smearing the speech	Large in a bare or tiled room, small in a carpeted one	No macOS setting compensates for the room; you have to change the room or the mic
Background noise	Competing sound lowers the ratio of your voice to everything else	Large in a cafe or open office, negligible in a quiet room	Voice Isolation is set per app, so Dictation cannot use it
Speaking style	Complete phrases give the model context to resolve similar-sounding words	Moderate, and it is the easiest habit to get wrong	Over-enunciating makes accuracy worse, not better
Language and region	Puts you on the right model for your accent, and often on the on-device path	A one-time fix, sometimes a big one	Only one of your languages is active at a time
Spoken punctuation	Removes the guesswork by naming the mark instead of hoping for it	Large for clean output, at a real cost to fluency	Saying every comma out loud breaks the flow of thinking aloud
Session handling	Avoids losing a sentence to the 30-second silence stop	Recovers lost text; changes no word-level errors	There is no setting to lengthen or remove the silence window
Custom vocabulary	The only way on macOS to teach a proper noun, up to 1,000 terms	Decisive for jargon, if your work is full of it	It is a Voice Control feature, and Voice Control blocks Dictation
A different speech model	Changes the thing generating the errors rather than the audio feeding it	The only lever that moves proper nouns and punctuation together	Costs money, and most good options need Apple Silicon

The microphone is the biggest lever, and distance is why

Get a microphone within a hand-span of your mouth and you will see a bigger improvement than any setting change. The reason is not that built-in Mac microphones are poor. It is that they sit in the laptop body, an arm's length away and pointing up into the room, and that distance changes the physics of what reaches them.

Speech loses energy as it crosses a room, so a distant microphone gets a quieter version of your voice sitting in the same amount of background noise. Worse, in any enclosed space the sound also bounces off walls, desks and screens and arrives again a few milliseconds late. Those late reflections pile on top of the direct sound. Speech researchers describe the result as the temporal smearing of the sub-band envelopes, and it is a known cause of degraded recognition accuracy. The consonant edges that tell "fifteen" from "fifty" are exactly what gets smeared.

This gap is big enough that the research field treats distant audio as its own problem. Recognizing speech captured away from the speaker needs a whole separate front end for dereverberation, source separation and beamforming, and a back end retrained for those conditions, which is why far-field recognition pipelines look nothing like close-talk ones. Your Mac is running the general path, not that specialist one. Moving the microphone to your mouth sidesteps the whole category of problem.

What to do, in order of effort:

Use whatever you already own. Wired earbuds with an inline mic, AirPods, a gaming headset. Anything mounted near your face beats the laptop.
Select it properly. System Settings, Sound, Input. Pick the device and watch the level meter move as you talk at your normal volume. A weak level is read as silence, which triggers early stops as well as errors.
Do not chase a studio microphone. A large desk condenser sits further away and hears more of the room than a cheap headset does. Proximity beats price here.
Skip Voice Isolation. Apple documents mic modes as something you set inside an app that supports mic modes, such as FaceTime. It is per-app, and Dictation is not an app with that control, so the widely repeated tip to switch it on for dictation does not apply.

How you speak, and the mistake most people make

Speak in complete phrases at your normal pace. The instinct when dictation misreads you is to slow down and separate every word, and that makes it worse. Modern speech models use the words around a sound to decide what the sound was, so an isolated word arrives with nothing to disambiguate it. Say "the discharge summary is ready" as one phrase and the model has four words of context for the medical sense of discharge. Say it one word at a time and it has none.

Two habits that do help. First, decide roughly what you are going to say before you press the key, because a half-formed sentence full of restarts gives the model contradictory context and gives you more to edit. Second, keep your volume level. Trailing off at the end of a sentence is where a surprising share of dropped words come from, since the final syllables land closest to the noise floor.

Punctuation, language and the settings worth changing

Two settings are worth a look, and one of them is the reason a lot of people think dictation cannot punctuate at all.

Language and region. In System Settings, Keyboard, Dictation, set the language to the variant you actually speak, not the one that came with the Mac. An Irish or Indian English speaker running US English is asking the wrong model to do the work. The setting also decides whether your dictation runs locally or reaches Apple's servers, since on-device support is a subset of the languages Dictation covers. If you move between languages, multilingual dictation on Mac covers how the switching works.

Punctuation. In supported languages Dictation inserts commas, periods and question marks automatically, and that can be turned off in Keyboard settings, so check which state yours is in before concluding it cannot do it. You can also name any mark out loud, and the spoken commands still exist for formatting: new line, new paragraph, caps on, all caps, numeral, no space, tab key. The trade is fluency. Saying every comma aloud interrupts the thinking you are trying to capture, which is why it suits short structured text better than a first draft. The full breakdown of both routes is in how to dictate punctuation on a Mac.

Session length. There is no setting here, only a habit. Dictation ends after 30 seconds without detected speech, so if you think in long pauses, expect to re-trigger it. Keeping each pass to one paragraph and restarting deliberately loses less text than discovering the microphone shut off two sentences ago. The full setup walk-through, if you want it from the beginning, is how to dictate on a Mac.

The lever almost nobody mentions: Voice Control's custom vocabulary

There is one way to teach macOS a word, and it is not in Dictation. Voice Control, the accessibility feature in System Settings under Accessibility, holds a custom vocabulary of up to 1,000 terms. You type the word as you want it entered, then say it so Voice Control learns your pronunciation, and you can import a whole list from a text file. Apple's examples are a CAD command and a clinical term, which tells you who it was built for.

For anyone whose errors are almost all proper nouns, this is the only built-in answer that exists. It comes with two honest catches. Voice Control and Dictation are not designed to run at the same time, so switching Voice Control on to gain a vocabulary means giving up the Dictation shortcut you were using. And the vocabulary option is dimmed in languages that do not support it, so the feature may not be available to you at all.

It is a genuine option and worth knowing about, especially for accessibility users who are already in Voice Control. It is a poor trade for someone who just wants better dictation, because you are swapping the whole input method to fix one class of error.

Where the levers run out

Once your microphone is at your mouth, your room is reasonable and your language is right, you have collected nearly everything the built-in feature has to give. What remains is fixed by the design, and no further tuning touches it.

Proper nouns and technical vocabulary stay wrong, because there is no vocabulary to add them to. The recognizer never improves with use, because there is no per-user adaptation, only an opt-in that helps Apple's next model. Punctuation stays at whatever the automatic behavior gives you unless you dictate the marks yourself. And the session still ends when the silence detector decides you have finished, which is the single most common complaint about the feature and the one thing no setting adjusts.

At that point the only remaining lever is the model. That is worth saying plainly rather than dressing up: if you have fixed the audio and the output is still not good enough for the work, you are asking a general-purpose recognizer for something it was not built to do, and the fix is a better recognizer. For casual notes and short messages, the built-in one is free and already installed, and stepping up is not worth it. For work where the transcript is the deliverable, it is.

This page is the tuning pass, the levers that live inside macOS. The replacement decision is the next page along: Apple Dictation alternatives, sorted by the limit you hit starts from the six limits no lever moves and says what removes each one, including which languages Apple leaves off its on-device list.

What macOS 26 changed, and what it did not

Apple did ship a much faster on-device speech engine, but it went to developers rather than into the Dictation panel. macOS 26 introduced new Speech framework APIs, SpeechAnalyzer and SpeechTranscriber, that any app can call to transcribe on-device. In hands-on testing at MacStories, a small tool built on them got through a 34-minute file in 45 seconds against 1 minute 41 for MacWhisper running Whisper Large V3 Turbo, with no noticeable difference in transcription quality. Both still tripped over proper nouns and camel-cased words.

Two things follow. Updating macOS does not hand you those speeds in the Dictation shortcut, because the new engine is an interface for apps, not a rework of the built-in feature. And the limits that annoy you most, no custom vocabulary and no personal adaptation, are not speed problems, so a faster engine does not address them. The interesting consequence is that third-party Mac dictation apps now have a genuinely good local option to build on, which is a point in favor of stepping away from the built-in path rather than a reason to stay on it.

Which apps improve Mac dictation accuracy for professional use?

For professional dictation the working options are apps that run a modern speech model in place of the built-in one, and they split by where the model runs. On the local side, Parakeety runs NVIDIA Parakeet TDT 0.6B v3 on the Apple Neural Engine as its only engine, with push-to-talk and automatic punctuation, for $25 once. MacWhisper and VoiceInk run local Whisper or Parakeet models and also handle recorded files. SuperWhisper offers a model picker that spans local and cloud, so what you get depends on your configuration. Talon Voice is a full hands-free control system rather than a dictation app, and the accessibility community relies on it.

On the cloud side, Wispr Flow and Aqua Voice pair a server model with post-processing that removes filler and tidies grammar, which is real value if you are happy for the audio to leave the Mac. Dragon, the tool lawyers and clinicians remember for its trained vocabularies, is no longer sold for the Mac at all.

The model itself is worth a sentence, because it is the part doing the work. Parakeet TDT 0.6B v3 is a 600-million-parameter FastConformer model covering 25 European languages with automatic punctuation and capitalization built in, and it reports 4.85% word error rate on English Fleurs on its own model card. The head-to-head against Whisper, including why the two behave differently in a pause, is in Parakeet versus Whisper. The wider survey of the field, weighing features beyond accuracy, is the best Mac dictation apps in 2026.

How I approach this

I make Parakeety, a one-person Mac app, so I have an obvious interest in you finding the built-in feature limiting. That is exactly why I have put the free levers first and been specific about what each one is worth. I built the app after running out of tuning options myself, which is the honest version of the story in why I built Parakeety. If a headset and the right language setting fix your dictation, you should stop there and keep your $25. Most people who are still unhappy after that are hitting the vocabulary and adaptation gaps, and those are the ones a different model actually solves.

FAQ

Why is Mac dictation so bad?: Mostly because you cannot teach it anything. The talk to text built into a Mac is Apple Dictation, and it has no custom vocabulary, so a client name or a drug name is whatever the general model guessed, and the custom vocabulary panel that does exist on macOS belongs to Voice Control rather than to Dictation. It does not adapt to your voice either: the Improve Siri and Dictation switch donates audio samples to Apple to improve its products generally, not to personalize your own transcripts. Add a built-in microphone sitting a laptop away from your mouth and pointing up into the room, and most of the errors people blame on the model are really coming from the setup around it.
How can I improve my dictation accuracy?: Fix the audio first, because it is the cheapest and largest gain. Get a microphone within a hand-span of your mouth, which raises the level of your direct voice against room reflections and background noise. Dictate in a room with soft surfaces rather than bare walls and glass. Speak in complete phrases at your normal pace instead of one exaggerated word at a time, because the model uses the surrounding words to decide what it heard. Then set your dictation language to match the way you actually speak. Only after all of that does changing the speech model itself become the thing that matters.
Which apps improve Mac dictation accuracy for professional use?: For professional work the practical options are apps that run a modern speech model in place of the built-in one. Parakeety runs NVIDIA Parakeet TDT v3 on the Apple Neural Engine with push-to-talk and automatic punctuation, for $25 once. MacWhisper and VoiceInk run local Whisper or Parakeet models and suit transcribing recordings as well as live dictation. SuperWhisper gives you a model picker spanning local and cloud options. Wispr Flow and Aqua Voice are cloud services with strong post-processing, but the audio leaves your Mac. Dragon is no longer sold for the Mac, so the trained custom-vocabulary workflow that lawyers and clinicians remember is not available there.
How do I improve my Mac microphone for dictation?: Move it closer and point it at your mouth. The built-in microphone sits in the laptop body, so it hears your voice after it has crossed the room and bounced off the walls, which is the condition speech researchers call far-field and treat as a separate, harder problem than close-talk audio. Any wired headset, a pair of earbuds or a desk microphone within a hand-span of your mouth removes most of that. Then check that the right input is selected in System Settings under Sound, and that the input level moves clearly when you talk. One thing that will not help: the Voice Isolation mic mode is set inside an app that supports it, such as FaceTime, so it is not a switch you can apply to Dictation.
Can I add custom words to Mac dictation?: Not to Dictation itself. Apple does provide a custom vocabulary of up to 1,000 terms, for things like CAD commands or medical words, but it is a Voice Control feature, so you reach it in System Settings under Accessibility and then Voice Control. The catch is that when Voice Control is on you cannot use Dictation, so you are choosing between the two rather than adding vocabulary to the dictation you already use. If proper nouns and technical terms are the bulk of your errors, changing the speech model is a more direct fix than working around the vocabulary gap.
What is the most accurate dictation software?: There is no single winner, and the honest reason is that accuracy depends more on your setup than on the software. Three things decide it: how close the microphone is to your mouth, whether the model has met the names and jargon you use, and how well it matches your accent. A modern local model such as Parakeet or a recent Whisper build will usually beat Apple Dictation on punctuation and proper nouns, but Apple publishes no accuracy figure for Dictation, so there is nothing official to compare it against. Published benchmark scores come from clean read-aloud audio, not from a laptop in a noisy room.

Try it

If you have already moved the microphone and fixed the language and the output still is not good enough, the remaining lever is the model. Parakeety runs Parakeet TDT v3 on the Apple Neural Engine, punctuates and capitalizes as it goes, and covers 25 European languages with automatic detection. Hold the section key, talk for as long as you like with all the pauses you need, release; your words paste at the cursor in whichever app you were typing into. There is no silence detector deciding you have finished, and audio never leaves the machine. It needs Apple Silicon and macOS 14 or later. There is a free 7-day trial with no card required. After that it is $25 once.

Try Parakeety free →