Dictation Mistakes and How to Make Voice Typing Accurate

Most dictation errors have five causes, and each one has a fix. Two of them, noise and distance from the microphone, degrade every word you say, so fix those first. The other three affect specific words: sound-alikes such as "flour" for "flower", names and jargon the engine has not learned, and quiet word endings such as the "-ed" in "walked." Vocabulary features and phrasing help with the names and the sound-alikes. Proofreading catches what remains, but it has to be done by meaning, because a dictated error is usually a real word that looks right on screen. A few changes to your words are not mistakes at all: profanity filters and number formatting alter the output on purpose.

Most dictation errors come from five predictable causes

The five causes differ in how often they hit and in how much you can change. The table sorts them by what you can do about each one.

Noise and distance are the two causes that act on every word. In a 2004 study of an oral-history archive with noisy, elderly and accented speech, researchers found that signal-to-noise ratio and syllable rate were the two dominant factors explaining the overall word error rate ¹. The study found no evidence that accent or speaker age affected recognition in that corpus ². That study used a 2004 recognizer, so its absolute error rates are not a guide to today's engines. The ranking of the factors is the useful part.

The reason the first two causes matter most is that they change the sound the engine receives. A better word list or a better sentence cannot recover a syllable that was lost in noise. For a fuller guide to the room, see dictation and background noise, and for the hardware, the best microphone for dictation.

Sound-alike words are the errors a spell checker never catches

A homophone is a real word, so a spell checker accepts it. "Their", "there" and "they're" all pass, and so do "to", "too" and "two." This is a common kind of dictation error that looks correct on screen, which is why homophones need their own check.

Apple's Dictation guide describes the problem directly. Ambiguous text is marked so you can correct it: "Ambiguous text is underlined in blue. For example, you may get the result “flour” when you intended the word “flower.”" ³ The underline is useful, but it only appears where the engine is unsure. A confident wrong choice gets no underline, so you still need to read the whole passage.

The engine chooses between sound-alikes using the words around them. That is the basic design of modern recognizers, and it is why the phrasing advice later in this guide matters. When you say "I will meet you over there," the surrounding words make "there" the likely choice. When a sound-alike appears alone, or after a pause that cuts the sentence in two, the engine has less to go on.

The fix is simple but takes discipline. Read each underlined word and each homophone and ask whether the meaning is right, not whether the spelling is right. The method is covered in how to proofread dictated text.

Names, brands and jargon need a custom vocabulary or a replacement rule

Names, product brands and technical terms are the errors that repeat. An engine that has never seen "Okonkwo" will often write two ordinary words that sound like it, and it will do so every time. Cloud speech services let developers supply a list of phrases, but the evidence that this helps is limited.

Google's model adaptation feature weights phrases you supply. Google's own guide says that by default the effect is relatively small: "By default model adaptation provides a relatively small effect, especially for one-word phrases." ⁴ The boost setting increases the weight of some phrases over others ⁵. Microsoft's Azure AI Speech phrase list also takes a weight, which you set within a range of 0.0 to 2.0 ⁶.

These are developer features for cloud services. They show how the feature is used in cloud services, not that your desktop app has a field for your surnames. Apple's Dictation guide, as read on 2026-10-16, describes choosing an alternative for an underlined word and typing or dictating the correct text, and it does not describe a custom word list (Apple Mac Help: Dictation). The Windows voice typing page lists settings such as Automatic punctuation, Filter profanity and Default microphone, and the part we read does not describe a custom word list (Microsoft Support).

If your app has no word list, use this sequence:

1. Dictate the name as you normally would, so the rest of the sentence stays intact.
2. Find every occurrence after you finish, with the find-and-replace tool in your editor.
3. If the app has a personal dictionary or an add-word option, add the name once so it is learned for later.
4. Say the name inside a full sentence. The surrounding words help the engine, as the next section explains.

The same approach works for jargon. Keep a short list of the terms you use most, and check them after each session until the errors stop.

Full sentences give the engine the context it needs to choose between sound-alikes

Speaking in complete clauses gives the engine the context it uses to pick between sound-alike words, and it is the most practical habit a dictator can build. We did not find a measured study that compares phrase-by-phrase speech with word-by-word speech on the same material, so treat this as the design of the systems and check it on your own draft. You can test it in five minutes: dictate the same paragraph once in short fragments and once in full sentences, then count the errors in each.

In practice this means four habits:

• Finish the clause before you pause. A pause after "I will meet you" and another after "over" cuts the sentence where the engine needs context.
• Do not stop after each word. Word-by-word speech sounds clear to you, but it removes the context the engine uses.
• Let the software place punctuation. The Automatic punctuation setting in Windows voice typing inserts punctuation from what you say ⁷ (Microsoft Support), and automatic punctuation works best when the sentences are complete.
• Pause at the end of each sentence, not inside it. A pause at the end of a sentence also marks where to stop and check it when you proofread.

Pauses matter for proofreading too, but in a different way. In online studies, adding a pause between words improved how well people identified speech recognition errors by ear, compared with no pause ⁸. That study measured listening, not recognition, so use the pauses when you review, and keep your dictation in full sentences. The broader habit of dictating without stopping to edit is covered in how to dictate.

Microphone distance and noise matter as much as how fast you speak

Speaking more slowly is a common instinct, but the evidence puts speaking rate on the same level as noise, not above it. In the 2004 MALACH study, syllable rate and signal-to-noise ratio were the two dominant factors ¹. Start with the microphone and the room, then adjust your pace.

The distance effect is large in controlled tests. A study that re-recorded read speech with a distant microphone found that, for distant speech, "the ASR performance is dramatically reduced" compared with close-talking input ⁹. The close-talking baseline in the same study was a word error rate of 3.7 percent for read sentences recorded in a studio ¹⁰. That is a best case, with read speech in a studio, so expect more errors in ordinary dictation. The distant result is a comparison of the same speech at two distances, and the direction is what matters.

Noise can also hide the quietest sounds, which are often word endings. When a passage with lost endings was recorded in a loud room, check the room first and the vocabulary second. Move closer to the microphone, close the window, and turn off the fan before you blame the engine. A noise-canceling microphone helps when the room cannot be quieted.

Connection matters too. Microsoft says that Windows voice typing "uses online speech recognition, which is powered by Azure Speech services" ¹¹, and that to use it "you'll need to be connected to the internet, have a working microphone, and have your cursor in a text box" ¹². If you dictate somewhere the connection is weak, check it first. For the offline question, see does dictation work without internet.

Some changes to your words are deliberate filters, not mistakes

Some changes to your text are made on purpose. The engine heard the word and the software replaced it. That is a filter, not a recognition error, and it can be turned off in many products.

Profanity filtering is the most common example, and the products differ:

The table shows three different designs. Azure masks by default, Google masks only when asked, and Windows lets you choose in the voice typing settings. A reader comparing products should not assume that one setting name means the same thing everywhere.

Number formatting is another deliberate change. Azure's output applies inverse text normalization, which "is conversion of spoken text to shorter forms," so that spoken numbers may be written as digits ¹⁹. You may want "two hundred" written out in a legal document, or "200" in a spreadsheet note. Check the output for both.

If you need an exact transcript, for example a verbatim record, look for the filter and formatting settings before you dictate. Turn off the filter where the product offers that choice, and proofread the numbers and the swear words, because a filter changes what you see but not what you said.

Proofreading dictated text needs a different method than proofreading typed text

Dictated errors are harder to catch than typed ones because the engine writes valid words. A typing mistake is often a broken spelling, which your eye catches. A dictation mistake is usually a real word in the wrong place, which your eye reads as correct. A spell checker, or a quick skim, will pass it.

The evidence on listening back is sobering. In an in-lab study, participants "missed identifying over 50% of speech recognition errors when listening to audio output of the recognized text" ²⁰. The same study found that identification gets worse at higher speech rates, and it tested 300 words per minute ²¹. Both results come from listening to synthetic read-back, not from reading a draft on screen. We did not find a measured detection rate for visual proofreading of dictated text, so do not assume the 50 percent figure applies to your own screen.

A method that fits the errors:

1. Read the whole passage once for meaning, not for spelling. Ask whether each sentence says what you meant.
2. Check every underlined word and every homophone, using the meaning test from the sound-alike section.
3. Check endings and small words: plurals, past tense, "a" and "an", "is" and "was", and "not".
4. Check names and numbers against your source.
5. Read aloud or listen with pauses between words, which the study above supports for listening.

The full process, including when to proofread, is in how to proofread dictated text. It is worth learning, because dictation errors are easy to miss and a reader will not catch them for you.

Accuracy you can expect today depends on the engine, the speaker and the room

There is no single accuracy figure for dictation, because the result depends on three things: the engine, the speaker and the room. Be careful with any single number, including vendor claims, unless you know the test conditions.

One controlled comparison of speech and keyboard entry is a 2018 study of short messages on phones. Speech produced fewer errors during entry than the keyboard, but left slightly more errors in the final text ²². The chart shows those numbers. It is a study of phones, short messages and a 2016 recognizer, so the gap today may be smaller, and the pattern is what to notice.

Speaker differences matter as well. In a 2020 study of five commercial ASR systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ²³. The authors traced the gap to the acoustic models used by the systems ²⁴. That study used interview audio, not dictation, and its systems are from 2020, so it does not predict your results. It does show that an engine can perform very differently for different speakers, which is why a single accuracy claim is not enough.

The room is the other variable, and the gains from noise work are real but modest in the 2004 study: noise compensation reduced the word error rate by 1.1 percentage points ²⁵. The bigger gains come from the microphone position and the quiet room, which you control directly.

Put together, the practical order is:

• Fix the room and the microphone. They affect every word, and they are the cheapest fixes.
• Speak in full sentences and let the engine use context.
• Add names and jargon to a word list where the app offers one, or fix them in review.
• Check the filters and formatting if the output looks altered on purpose.
• Proofread by meaning, with a second pass for sound-alikes and endings.

If you want to try these fixes on a Mac or Windows PC, Yap, which we make, is an on-device dictation app for both. The free built-in tools described in how to turn on voice typing may be enough, so start there. For a comparison of accuracy across human and automatic transcription, see human vs AI transcription.

The rest of the dictation guide covers the steps that follow

Continue with the rest of the dictation guide, starting with dictation punctuation for spoken punctuation, dictation commands for editing by voice, and dictation not working if the feature stops. For the questions this guide answers in short form, see how to improve dictation accuracy, why talk to text censors words and how to make dictation spell names correctly. Terms used here are defined in the glossary: dictation, voice typing, spoken punctuation.