Dictation in noisy office Mac: what error rates look like in 2026

Dictation in a noisy office on Mac usually gets less accurate as background talk gets louder, and Apple does not publish an error rate for its built-in Dictation. The best public numbers come from a 2024 study of Whisper, which found a 44% word error rate for the large model in pub noise at 0 dB.

Those numbers are useful, but they are easy to misread. This guide explains what an error rate measures, what the published figures cover, and what they leave out. It also shows how to test your own desk, and which fixes move the result most.

What an error rate means for Mac dictation

Speech researchers usually report word error rate, or WER. WER counts every word the system gets wrong, then divides by the number of words you actually said. A lower number means less fixing for you.

WER counts three kinds of mistake:

• A substitution is a wrong word in place of the right one, such as “desk” for “deck.”
• A deletion is a word you said that never appears in the output.
• An insertion is a word that shows up in the output even though you never said it.

WER is a good rough guide, but it misses what hurts in an office. A wrong name, a wrong figure, or a flipped “not” costs you far more than a dropped “the.” Scoring setups also differ on punctuation and capitalization. Check how a number was calculated before you compare it with another one.

Does Apple publish an error rate for Dictation?

No. Apple's Dictation guide and troubleshooting page cover how to start Dictation, where the settings are, and what limits apply. They do not give accuracy figures for any noise level. That gap is why most people end up guessing.

Here is what Apple does say. Dictation has no length limit. It stops by itself when it detects no speech for 30 seconds. You can stop it with Escape, the Microphone key, or the Dictation shortcut. For noisy rooms, Apple's troubleshooting page says to avoid background noise. It adds that in a noisy room or one with a lot of echo, a headset microphone may help.

Apple also says you may need an internet connection for Dictation. The text below Dictation in Keyboard settings shows whether general text dictation is processed on the device rather than sent to Siri servers. Check that line on your own Mac. Do not take the answer from a blog post, including this one.

Apple's Dictation guide uses macOS 27 steps, and it adds one more change. With Apple Intelligence, an on-device model improves spelling, punctuation and capitalization. Apple says this works in English on Macs with an M3 chip or later and at least 12GB of unified memory. If your Mac is older or runs an earlier macOS, the menus and features may differ.

Of these, the microphone source is the setting most worth checking first.

What the published numbers show for noisy rooms

The most useful public study I found is Patman and Chodroff's 2024 paper, “Speech recognition in adverse conditions by humans and machines,” in JASA Express Letters. The authors tested British English listeners and four speech recognition systems: wav2vec 2.0 base and large, and Whisper base and large. They played two noise types, speech-shaped noise and pub noise, at 0 dB and 8 dB signal-to-noise ratio. Some runs also used cotton face masks.

Signal-to-noise ratio, or SNR, compares how loud your voice is with how loud the background is. At 0 dB, the noise is as loud as the speech. At 8 dB, your voice is clearly louder. Your office could land anywhere in that range, depending on your mic and where you sit.

The headline result is for Whisper large. Its mean word error rate in pub noise at 0 dB was 44%. The authors compare that with the original Whisper paper by Radford et al. (2023), which they cite at roughly 18% for pub noise and roughly 17% for white noise at 0 dB. Those figures are estimates read from a graph, not numbers stated in the text, so treat them as approximate.

The authors also say that humans kept an advantage over every system. Whisper large beat the human listeners in every condition except pub noise. In that one condition, their model put Whisper large on par with the humans. The model uses a different measure than raw WER, so do not compare those results directly with the 44% figure.

Two limits matter most. The study tested Whisper and wav2vec 2.0, not Apple's Dictation engine. It also did not test office noise. Pub noise is the closest match to office chatter, but it is still a different sound. Use these figures to understand the shape of the problem. Do not use them to predict your own score.

Why a shared office is harder than a test room

A test room controls the noise, the distance and the microphone. Your desk controls none of those. Four things make an office harder than the study.

Nearby talk comes first. Pub noise sounds like many people talking at once. Colleagues on calls are often the hardest part, because their voices sit in the same range as yours.

Second, you change how you speak. People often raise their voice when a room gets loud. Louder speech is not always clearer speech, and a mic can pick up the strain as much as the words.

Third, hard surfaces add echo. Glass walls, bare desks and high ceilings bounce sound back into the mic. Apple names echo as a reason to try a headset.

Fourth, your mic sits in a poor spot. A laptop mic on a crowded desk picks up keyboard clatter, fans and talk from across the room. It also sits far from your mouth, which lowers your voice relative to the noise.

The 30-second stop adds its own problem. If you pause to listen to an interruption and stay quiet for longer than 30 seconds, Dictation ends. Restart it with the Microphone key or the Dictation shortcut, then check the last sentence before you continue.

Fixes that lower your error rate, most important first

The microphone does most of the work. Settings come second, and habits come third. This is the order I would try them in.

1. Put a headset mic close to your mouth. Apple recommends a headset in noisy or echoey rooms. A closer mic raises your voice above the background, which addresses the SNR problem directly.
2. Pick the mic on purpose. Automatic uses the device you are most likely to use, and that may not be your headset. Go to System Settings > Keyboard > Dictation, open the Microphone source menu, and choose the headset.
3. Sit away from hard surfaces. If you can, move away from glass and bare walls, since they reflect your voice back into the mic.
4. Keep your volume and distance steady. Speak at your normal volume and keep the mic at the same distance. A change in distance changes the level, and the level changes the errors.
5. Use the on-device enhancement if you qualify. Check the text below Dictation in Keyboard settings. You need an M3 chip or later, at least 12GB of memory, macOS 27, and English.
6. Teach the engine your names. Names and project terms are where office errors stand out most. Keep a short list and check it before you send anything. The names and jargon guide for Mac dictation walks through custom vocabulary.

How to measure your own error rate in the office

The only test that tells you about your room is a test in your room. It takes a few minutes and one passage.

1. Choose a passage you already know. Pick a paragraph with a name, a figure and a project term. Keep the original text open in another window.
2. Dictate it at your usual desk. Use the mic you use every day, and start Dictation from Edit > Start Dictation or the Microphone key.
3. Compare the output with the original. Go word by word and mark each substitution, deletion and insertion.
4. Divide the wrong words by the words in the original. The result is the WER for that run.
5. Repeat the test while a colleague talks nearby. Compare the two results, and note which words broke.

Say the original line is “Please send the deck to Priya before lunch.” Dictation returns “Please send the desk to Maria before lunch.” Both “desk” and “Maria” count as errors. The name costs you more time than the count suggests, because you have to check it before anything goes out.

Decide now how you will treat punctuation, and apply that rule to every run. Mixing rules makes two runs impossible to compare. If you want a faster way to score pasted text, try the dictation accuracy checker.

Common mistakes that make dictation look worse than it is

• Testing once in a quiet room. A quiet room tells you about the quiet room. Measure at your desk.
• Judging the engine before the mic. A wrong microphone source can make any engine look bad. Check the setting first.
• Reading 44% as Apple's number. That figure belongs to Whisper large, in one test with pub noise at 0 dB.
• Scoring only the word count. One wrong name or figure can matter more than several small wrong words.
• Dictating a whole email in one go. Dictate in chunks, check each one, and fix names before you move on.
• Turning your head away from the mic. Speaking to a second monitor or a colleague moves your voice off the mic.

When to type instead of dictating in an office

Dictation is not the right tool for every message in a shared room. Some work goes better with your hands on the keys.

Type short replies to colleagues, especially when they need exact wording. Type passwords, URLs, code and figures. A wrong character in a URL is easy to miss after you have spoken it.

Dictate long drafts in a quiet spot, then edit them. If you write in Word, the Dictation in Word steps for Mac cover the menus for Microsoft 365 and Word 2024.

If your team sits in an open room, move long dictation sessions to a meeting room. A closed room cuts the background talk, so the same passage tends to come out cleaner there.

Where a voice assistant fits in a noisy office

Lazur dictates clean, formatted text at the cursor in any app when you hold fn. It also keeps snippets, a personal dictionary and clipboard slots. The personal dictionary is the part to test first, because names are the errors you will notice first.

It listens and looks at the screen only while you hold the shortcut, which suits a shared room. Hold ⌥ ⌘ to ask about something on screen, and the answer comes out loud. In an open office, you may prefer the answer as text instead.

This page gives no error rate for Lazur, so run the same passage test on it before you rely on it. The Apple Dictation comparison covers when the built-in tool is enough. If you want to try it at your desk, get the Mac app.