Most advice about bad dictation accuracy on a Mac is the same advice: speak clearly, don't mumble, buy a better microphone. It isn't wrong, it's just useless — it treats one symptom as one problem when there are several, and they have different fixes. The person whose Bluetooth headset is quietly degrading the audio and the person whose recognizer has never heard their colleague's surname both say "voice to text keeps getting words wrong," and neither is helped by enunciating harder.

Diagnose by the shape of the error. What comes out wrong tells you which layer broke.

What you're seeing Usually means
Whole phrases garbled or dropped, worse on calls The mic switched to a Bluetooth call profile
Fine at your desk, bad in a meeting room Distance and reverb
One specific name or acronym wrong every single time No custom vocabulary
Technical terms wrong, plain English fine The engine is the ceiling
Plausible words in the wrong language Language setting vs. what you actually spoke
Words right, sentences unreadable Punctuation and formatting, not recognition
Dictation stops on its own A silence timeout, not accuracy at all

When Mac dictation is not working well, work down that list in order. The first two account for most of it, and both are free to fix.

1. Your microphone sets the ceiling

Nothing downstream can fix audio that was already degraded, and on a Mac the most common cause is a wireless headset.

Classic Bluetooth has to switch from its high-quality playback profile to the hands-free call profile the moment an app opens the microphone. The mic path in that profile is mono and heavily compressed — narrowband at 8 kHz, wideband at 16 kHz — and you can hear the switch happen, because your music turns to mud at the same instant. Speech models work from 16 kHz audio, so a compressed narrowband signal is genuinely less to work with. This isn't a specific brand's failing; it's how the profile works, and it applies to most Bluetooth headsets.

There's a real exception worth knowing, because a lot of advice on this hasn't caught up. Apple's studio-quality audio recording covers AirPods 4, AirPods 4 with ANC, and AirPods Pro 2 on macOS Tahoe 26, and it works on the Mac, not just iPhone. It's a substantial improvement over the old call-profile behavior. It doesn't apply everywhere automatically — Apple lists specific apps and app categories that support it, so whether your dictation app gets that quality depends on the app.

The reliable fix is the same. Open your Mac microphone settings in System Settings → Sound → Input and pick your built-in microphone from the device list — macOS names it after your model, like "MacBook Pro Microphone." Connecting a headset switches your input to it, and nothing tells you that happened. On laptops and iMacs, the built-in option is a three-mic array with directional beamforming, and for recognition purposes it beats most wireless headsets. Keep the earbuds for output if you like; just don't let them be the input.

If you want a better dictation microphone for your Mac than the built-in one, the upgrade is a wired USB microphone, not a nicer wireless one. And if you're unsure whether audio is your problem, record thirty seconds in Voice Memos with your current setup and listen back. Muffled or boxy to your ears is worse to the model.

2. Distance and reverb, not volume

The second thing that destroys accuracy is the room. Every hard surface sends a delayed copy of your voice back into the mic, and that smears the acoustic boundaries between words — exactly the information the recognizer uses to tell them apart.

Practical version: sit 20–30 cm from the mic rather than an arm's length away, and kill the obvious noise sources — the desk fan, the open window onto traffic, the mechanical keyboard you're typing on while you talk.

Meetings are where this bites hardest, and there's a second effect on top of reverb. If you're on speakers instead of headphones, remote audio loops back into your microphone. Anything transcribing the meeting locally hears their words on your mic channel too and attributes them to you. Headphones solve it outright. Software can also clean it up after the fact — Vext compares the mic and system streams against a shared meeting clock and drops mic lines that repeat what the system stream already captured within a two-second window — but the physical fix is still the better one.

3. The engine is the biggest single step up

If the audio is clean and speech-to-text accuracy is still mediocre, you've hit the ceiling of the engine.

Apple's built-in dictation is designed to be instant, free, and always available on a machine doing a hundred other things. It has improved a lot, and for a quick reply it's fine — we covered where it's enough and where it stops in detail. But it's a different class of model from what a dedicated app runs. Heavier local engines on Apple Silicon — NVIDIA's Parakeet, the Whisper family — spend more compute per second of audio, and buy accuracy with it.

Cloud transcription APIs are the other lever, and they are a real option, but they mean your audio leaves the machine. For client calls, patient notes, interviews, or anything under an NDA, that alone rules them out — and the accuracy they buy is usually small on clean audio, larger only on genuinely hard recordings. On an M-series Mac, local-first is the default worth starting from.

Two things to know before you go chasing model size:

  • On clean speech, the gap between good engines is small. Clean audio plus ordinary vocabulary means a bigger model may change almost nothing.
  • The gap widens exactly where you're probably struggling: noise, accents, jargon, overlapping speakers.

So test with your own hard audio rather than trusting a benchmark. Dictate a paragraph containing the terms you actually use — product names, library names, the acronym your team says forty times a day — and compare engines on that. If you want a baseline before changing anything, our dictation accuracy test scores a read passage word by word in the browser — the technical-terms and proper-names passages show you fastest which category is failing you.

4. Proper nouns aren't an accuracy problem. They're a configuration problem

This is the one people wait longest to fix, and it's the cheapest.

No engine reliably produces a word it doesn't know exists. Your colleague's surname, your company name, an internal codename, kubectl — these come out wrong not because recognition is bad, but because the model is choosing the most probable known word that matches the sound. It will keep doing that forever. Speaking more clearly doesn't help, because the right spelling was never a candidate.

Apple gives you little here. Dictation itself has no custom vocabulary. Voice Control does — Apple's own page for it says you "can add words and phrases—individually or by importing a list" — but that's a separate accessibility feature for controlling your Mac by voice, available in some languages, not a vocabulary for dictating into a document.

Dedicated apps solve it with a vocabulary or replacement list. In Vext 1.3.0 the rules are pairs — what the recognizer produces, what it should have produced — applied to the raw transcript and then passed into the cleanup prompt as a preserve instruction, so the AI pass can't paraphrase your corrected spelling back to wrong. On engines that accept a hint up front (Apple's recognizer, Whisper-compatible endpoints) the terms are also fed in before transcription. That preserve step matters more than it sounds: a vocabulary feature that fixes the transcript and then lets a language model rewrite it has fixed nothing.

For one week, note every word you correct by hand. It'll come to ten or fifteen entries, mostly names, and that list removes most of your remaining editing.

5. Language settings and mixed speech

If you're getting plausible words in the wrong language, or your second language degrades badly mid-sentence, this is a settings problem rather than a model failure.

Monolingual recognition assumes one language and decodes accordingly, so an English product name inside a Spanish sentence gets mapped to the nearest Spanish-sounding thing. To change the dictation language on a Mac, open Keyboard settings, where dictation languages live: click Edit next to Languages to add or remove them, and once more than one is enabled you can switch mid-dictation from the language shown next to the cursor. Apps built on multilingual models often detect the language instead of asking.

Code-switching inside a single sentence is the hardest case for any engine, and the honest advice is to put the proper nouns in your vocabulary list so at least those survive.

6. Punctuation and formatting aren't recognition

A transcript where every word is right and the whole thing is one unpunctuated run-on isn't an accuracy failure. Recognition worked. What's missing is the pass after it.

You can dictate punctuation by voice ("comma", "new paragraph"), and auto-punctuation infers some from cadence, more reliably in short sentences than long ones. The better answer is to stop doing it by voice: speak naturally and let a cleanup pass handle punctuation, capitalization, and filler words afterward. On Apple Silicon that pass can run entirely on-device through a local model, so it costs you nothing in privacy. It's also what lets the same spoken paragraph come out as an email, a chat message, or bullets — the reshaping happens after recognition, not during it.

7. The "30-second limit" that isn't what you've read

Worth correcting, because it's repeated everywhere as a hard cap on how long you can dictate. Apple's own documentation says the opposite: "You can dictate text of any length without a timeout," and separately, "Dictation stops automatically when no speech is detected for 30 seconds."

That's a silence timeout, not a length limit. It ends the session when you've stopped talking for half a minute — which is generous if you're dictating continuously, and irritating if you compose by speaking, thinking for a while, then speaking again. If that's you, what you want is push-to-talk (hold a key while speaking) or a toggle hotkey (press to start, press to stop), so the session ends when you say it does rather than on a timer. Both patterns are covered in the Mac voice transcription setup guide.

The 60-second checklist

In order, highest impact first:

  1. System Settings → Sound → Input — select your built-in microphone explicitly, not a connected Bluetooth headset.
  2. Move to 20–30 cm from the mic. Turn off the fan.
  3. Wear headphones on calls so remote audio doesn't loop back into your mic.
  4. Record 30 seconds in Voice Memos and listen. If it sounds bad to you, fix that before blaming the software.
  5. Write down the ten words you correct most, and put them in a custom vocabulary.
  6. Check your dictation language, and add a second one if you switch languages.
  7. Stop dictating punctuation; use a cleanup pass instead.
  8. Still short on technical vocabulary? Try a different engine with your own hard audio — a local one first, since a cloud endpoint means uploading everything you say.

Six of those eight are free and take under a minute.

When it's actually the tool

If you've worked through the list and it's still not good enough, the tool is the constraint — usually because it has no vocabulary control, no choice of engine, and no cleanup pass, which means you've run out of things to tune.

That's the case Vext is built for, and it starts from the privacy side rather than bolting it on: speech recognition, AI cleanup, and meeting transcription all run on your Mac, so no audio and no transcript leaves it. On top of that, a choice of engine (Parakeet or Apple's recognizer locally, with an OpenAI-compatible endpoint available if you explicitly want one), a custom vocabulary that survives the AI pass, dictation modes that shape text for where it's going, and meeting transcription that handles the echo problem.

No cloud by default, no account, no subscription and no locked tier — $49 one-time, every feature included.

brew install muvon/tap/vext

Start with the microphone, though. It's free, it takes ten seconds, and more people need that fix than need a different app.