Fix Mac Dictation Accuracy: Settings, Mic, and Vocabulary

Use one test sentence so you can tell what changed

Accuracy work goes wrong when you change three settings at once and judge the result on whatever you happened to say. Write one test sentence, keep it in a note, and reuse it after every change. Twenty-five words is enough if the sentence contains the things that break.

• A colleague’s name and a client or company name.
• A product name, repo name, or acronym you say weekly.
• A number with digits, plus one email address.
• One sentence that ends in a question mark.

Say it into the app you actually write in, not into a text editor you never use. Count corrections and write the number next to the change you made. Three data points beat a hunch, and the log tells you when to stop fiddling.

You will get better at reading your own test sentence, which does not make the engine better. That is fine. You are comparing configurations against each other, not measuring how you speak on a normal Tuesday.

Microphone distance and the room around it

Distance is the biggest lever you have, and it costs nothing. Aim for a stable position you can hold for a full paragraph without leaning in, roughly 30 to 50 cm from a laptop mic. A mic you only reach when you sit forward will fail halfway through a long email.

1. Open System Settings > Sound > Input and pick the device you think you are using.
2. Speak your test sentence and watch the level meter at your normal volume.
3. If the meter barely moves, get closer before you raise input gain.
4. If it pins to the top, back off or lower gain so consonants stop clipping.

Then clear the path. Laptop fans, an air vent, a desk fan, and a mechanical keyboard all land in the same microphone. Speaking across the mic rather than at it adds breath pops that come out as dropped syllables.

One habit fixes a common complaint: wait half a second after you press the shortcut before you start talking. Most reports of a missing first word are a take that started before the mic did.

Bluetooth caveats worth knowing before you blame the model

Bluetooth headsets sound fine on calls and often worse for dictation. When the headset mic engages, macOS changes the input route and the mic path runs at a lower quality than the music you hear through the same earbuds. Blurred consonants and a swallowed word after each pause are the usual signs.

Bluetooth still wins when you are walking or sitting in a loud room, because getting a mic near your mouth matters more than the codec. In a quiet office, the built-in mic is usually the better default.

Watch for the device switching itself. Video call apps reassign the input, and putting one earbud back in its case mid-session can move the mic. If accuracy fell off a cliff this week, check the input device before you download anything.

Settings that cause steady, repeatable errors

Setting problems produce the same failure every time instead of random misses. If one class of word fails on every attempt, look at language and permissions before you buy hardware or swap models.

• The dictation language matches the language you are speaking right now.
• The engine finished downloading rather than sitting half cached.
• Microphone permission is granted to the app you are dictating from.
• Accessibility permission is granted if the app inserts text at your cursor.
• Auri Voice needs macOS 15 or later, so confirm the OS before troubleshooting further.

Missing Accessibility permission looks exactly like an accuracy problem. Recognition works, no text appears, and you assume the model failed. Test in Notes and in Slack, since focus behavior differs between native and Electron apps.

Separate the two layers while you test. Auri Voice can keep speech to text and text improvement on your Mac, and the improvement pass is on: it repairs self-corrections, adds punctuation, formats lists, and removes filler. AI Processing can also use Auri Cloud or a bring-your-own-key provider if you choose. Misheard words are a recognition problem. Rewording is the cleanup pass doing its job.

Custom words for names, brands, and jargon

Custom vocabulary fixes what no setting can reach: surnames, client names, internal acronyms, product codes. Build the list from evidence rather than imagination. Scroll your sent mail and Slack from the past two weeks and pull the terms you had to retype.

1. Collect fifteen to twenty-five terms you correct most often.
2. Add them as the form you say out loud, not only the way you write them.
3. Dictate one real email and one real Slack update that use those terms.
4. Delete entries that never fire and keep the ones that cut edits.

In Auri Voice, custom vocabulary works with the Parakeet engines. If names drive most of your corrections, keep Fast English or Fast Multilingual as your daily pick and treat Whisper Large Turbo as the option for noise and accent problems that survive a good word list.

Refresh the list when a project starts and prune it when one ends. Twenty current terms help more than a hundred names from a client you stopped working with last spring.

Quiet room versus noisy room tactics

Noise strategy is distance first, engine second. Getting the mic closer to your mouth removes more error than any model swap, which is why a wrist or a headset beats a laptop lid across the desk in a loud space.

Auri Voice can use an Apple Watch as a close mic, which helps in loud environments, on a commute, when you need to whisper, or when the Mac is docked across the room. Audio is encrypted in transit and transcribed locally on the Mac, and the text lands on the laptop while it is open.

In loud places, shorten takes to one thought each. A thirty-second take in a cafe gives the engine more chances to guess wrong in a row, and one bad guess drags the rest of the sentence with it.

Change engines last, and judge over a week

Swapping engines is the most satisfying step and the least reliable one, because a new model resets your sense of what was broken. Do it after signal, settings, and vocabulary, or you will not know what actually helped.

1. Baseline with Fast English (Parakeet TDT v2) or Fast Multilingual (Parakeet TDT v3), plus your custom vocabulary, on the test sentence and one real email.
2. Run the same two texts on Whisper Large Turbo in the same room, at the same distance.
3. Try Whisper Base if you need a lighter local download, or Auri Cloud / Apple on-device if you want no download.
4. Keep the winner as your default for a full week before you decide.

Larger models cost disk, memory, and heat. One that looks excellent in a morning demo can feel slower on a warm MacBook at four in the afternoon with twenty tabs open, so judge it during a normal afternoon rather than a quiet test.

Auri Voice on macOS 15 or later switches among five local engines, Auri Cloud, and Apple on-device speech without a reinstall. Stop tuning when correction time drops far enough that you keep dictating tomorrow. Speaking runs around five times faster than typing, and that advantage disappears if you re-edit every line.

FAQ