Speech Recognition and Accents: Accuracy and Fixes

Speech recognition still makes more errors on accented English than on standard US English, and the gap is measurable. In a 2023 test on conversations recorded across many English accents, the best model averaged a word error rate of 19.7 percent ¹. On US English read speech, the same model had 2.7 percent ². That is about 7.3 times the error rate ³. The two tests differ in more than accent, so read the ratio as the size of the gap, not its single cause.

The gap appears to come largely from the speech a system was trained on, which you cannot change. The fixes you control work differently: they change the settings and the microphone, not the system. Pick the exact regional setting, use a multilingual model when you mix languages, teach the app your own words, and improve the microphone. Each section below gives the numbers behind one point and says where the evidence is thin. Every number on this page carries a reference that opens the exact sentence from its source, so you can check the figure yourself. For the broader research, see our guide to speech recognition accuracy. This page is part of our guide to dictation languages.

Accented speech is recognized less accurately, and read speech shows a smaller gap

Accented speech is recognized less accurately, and the drop shows up across systems. The EdAcc authors tested public and commercial English models on conversations between friends. They report a drop in performance across all models for speakers of Indian, Jamaican and Nigerian English ⁴. If you want a fuller answer for one accent, our question on speech to text for an Indian accent covers that case.

Read speech shows a smaller gap. On read English sentences from non-native speakers in the L2-ARCTIC corpus, Whisper reached a mean match error rate of 0.054 ⁵. The author describes that as approaching human-level accuracy. Spontaneous speech is harder, and the best system changes. RevAI had the best result on spontaneous speech, at 0.063 ⁶. Match error rate is a related measure, not the same as word error rate, so these figures cannot be compared directly with the EdAcc ones.

The table shows that speech type changes the result, not only the accent. Read speech gives the better result in both tests. The EdAcc rows show the largest difference, with conversation far more error-prone than read speech. The two L2-ARCTIC figures are much closer, and they use a different measure.

An average hides the spread. A figure of 19.7 percent describes the whole test set, and some speakers in it will have much lower error rates while others have much higher ones. That can help explain why the same app seems fine to one colleague and poor to another. When you judge a tool, use your own speech, not someone else's experience of it. The question on why dictation does not understand your accent covers the common causes one at a time.

Older systems struggled with dialect speech. In the 1996 and 1997 NIST CallHome evaluations of Egyptian Colloquial Arabic, the best word error rate was 61 percent ⁷. That test and today's are not comparable, so it shows the direction of the change, not an exact gain.

Sources point to the speech a system was trained on

The sources here point to the data a system learns from, not to the speaker. In a 2020 study of five commercial systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ⁸. That study compares speaker groups defined by race, not foreign accents. It still points to the same mechanism. The authors trace the gap to the acoustic models that the systems use ⁹.

Google gives a related explanation. It states that automatic speech recognition systems are most often trained from "typical" speech ¹⁰. If a system has heard little speech like yours, it is likely to make more errors on yours. That is also why why your dictation does not understand your accent has a training answer rather than a speaking answer.

This also explains why two apps can disagree about the same speech. Each app runs its own recognizer, trained on its own data, so one may handle your speech better than another. The Koenecke result covers a family of commercial systems, and the EdAcc result covers several public and commercial models, so neither depends on a single product.

The training effect is measurable across languages. In Whisper's tests, the log of the word error rate has a squared correlation of 0.83 with the log of the amount of training data for each language ¹¹. The chart shows the same pattern for four languages on the FLEURS benchmark.

The chart uses the FLEURS results for English at 4.2 percent ¹², French at 8.3 percent ¹³, Hindi at 21.5 percent ¹⁴ and Yoruba at 94.8 percent ¹⁵. Languages with less training data show higher error rates. Accents appear to follow the same logic: speech that resembles less of the training data gets more errors.

Whisper and AssemblyAI led read speech, but no single system won every test

No single system won every test. On read non-native English, Whisper and AssemblyAI were the top two. On spontaneous speech, RevAI led. The multilingual dictation guide compares what each system handles.

Older systems struggled with dialect speech, as the 1996 and 1997 Arabic result shows. The read-speech results above are better than the free-speech results in both tests. Recognition of free speech is still less reliable, so do not assume the best read-speech score applies to your own dictation.

The case for a multilingual model is usually strongest when you switch languages. The studies above do not isolate that effect for accents, so they do not show that a multilingual model is more accurate for your accent than a single-language model. If you mainly dictate in English, test both on your own voice before you choose. The guide to Whisper's supported languages lists which languages the model covers and how accurate each one is.

Choosing the regional variant closest to your speech is the cheapest test to run

Set the regional variant closest to your speech, because the setting selects which recognition model the app uses. No study in our sources measures this effect directly, so treat it as the cheapest test you can run, not a guaranteed gain. Whether to set US or UK English is a common version of this choice.

The choice of variants is wide and differs by platform. Apple's macOS feature list has 63 language and regional variants for Dictation ¹⁶. Microsoft lists 15 languages and dialects for Voice Access in Windows 11, including six English variants ¹⁷. Lists differ, so check the one for your own device. The guide to dictation languages compares them side by side.

On a Mac, the setting is in Keyboard settings, as Apple's Dictation guide describes.

1. Open System Settings and choose Keyboard.
2. Under Dictation, click Edit next to Languages ¹⁸.
3. Select the language and region closest to your accent.

If none of the variants matches your speech closely, choose the nearest one and then test the second-nearest, using the same passage for each. Two variants that look close on a list can score very differently in practice, so the test is the only reliable guide.

Apple says you can set up several languages and switch between them ¹⁹. It describes switching between languages, not recognizing two languages inside one sentence. A 2025 preprint on code-switching recognizers suggests why. Its authors say that spontaneous switching within a sentence causes language confusion, and that accent bias blurs the sound boundaries between languages ²⁰. If you mix languages in one sentence, read the guide to multilingual dictation and the glossary entry on code-switching first.

Custom vocabulary and a better microphone are the fixes you can control

Custom words help with the errors that come from names, product terms and words a system hears in the wrong way. Start with the words you correct most often in your own drafts, such as your name, your company, technical terms and words from your field. Keep the list short, because each added term is one more thing to check. Google's model adaptation feature lets you give some phrases more weight than others ²¹. By default, though, its effect is relatively small, especially for one-word phrases ²². Add a term, then test it, rather than assuming the setting worked.

Vocabulary tools differ by platform. On a Mac, Voice Control can import custom vocabulary terms from a text file with one term on each line ²³. In Windows Voice Access, the Add to vocabulary command is listed for improving recognition of complex or hard-to-pronounce words ²⁴. Check that the tool supports custom vocabulary in your language before you build a list.

The microphone matters as much as the words. Apple advises that a headset microphone may help when Dictation is used in a noisy room or one with excessive echo ²⁵. In a test with the same Whisper model, the single distant microphone recording of AMI meetings had about 2.15 times the word error rate of the headset recording ²⁶. If your accent is harder to recognize, a closer microphone removes one source of error that you can control.

Training on your own speech is a different kind of fix. On Windows, Microsoft describes Voice Isolation as needing a quick one-time setup so that Voice access can learn to recognize your voice ²⁷. It is the closest built-in step to training on your speech in the Microsoft documentation we read, and it is worth trying before you give up on a tool. In a 2019 Google study, personalization reduced word error rate for accented speech by 35 percent relative ²⁸. An earlier 2007 study of foreign-accented English found that the best combination of pronunciation modelling and acoustic adaptation cut word error rate by 46 to 71 percent relative ²⁹. These are research methods, not buttons in most dictation apps. Our question on training dictation to your accent explains what your app can and cannot learn.

Regional dialects also produce high error rates

In the studies we found, regional dialects also produced high error rates. A dialect differs in vocabulary and grammar as well as sound, not only in pronunciation. The 1996 and 1997 Arabic result is the clearest example: 61 percent word error rate on Egyptian Colloquial Arabic ⁷. The test was a different one from the 2023 English test ¹, so the two numbers only show that dialect speech is a hard case.

Native English varieties are also not all equal. In the EdAcc study, every model dropped in accuracy on Indian, Jamaican and Nigerian English ⁴. These are English varieties with their own pronunciation, so the regional setting can matter even when English is your first language. This is also why the term for non-native speech does not cover every case: a regional variety is not always a second language. Rely on your own test rather than a single benchmark.

For dialect speech, one practical fix is the custom word list. Regional words, local place names and everyday phrases from your area are terms a general model may miss. Add them to the list and test them, then use the same passage method from the test section to see whether the change helped. Dictating a few sentences in your own dialect before a longer job is a quick way to find the gaps. For a language with its own voice typing options, the guide to Hindi voice typing shows how the choices differ.

Test any dictation tool on your own voice before you pay for it

Test the tool on your own voice, with a passage you already know. A benchmark average tells you about a test set, not about you. The standard measure is the word error rate, which the 2020 study defines with the total number of words in the correct transcript as its denominator ³⁰. You can calculate it yourself in a few minutes.

1. Choose a passage of about 200 words that you would normally dictate, and keep the text in front of you as the reference.
2. Dictate the passage into the app you use most, at your normal pace and with the microphone you use every day.
3. Count the words the app gets wrong. Include substitutions, missing words and extra words.
4. Divide the number of errors by the number of words in the passage. The result is your word error rate.
5. Repeat the test with a headset microphone, then compare the two results.

The word error rate guide explains the formula and how to read it. Use the same passage for each app, so the comparison is fair.

Yap, which we make, is an on-device dictation app for Mac and Windows. You can run the same passage through it and through the built-in option on your device, and compare the results. For prices and downloads, see pricing and the download page.

Where to go next

If you want a short plan, do these in order. First, set the regional language closest to your speech and dictate a test passage. Second, add the words you correct most often, and test again. Third, repeat the test with a headset microphone in the room where you usually work. Fourth, keep the results with the tool you chose, so you can compare again after an update, since language support and accuracy change over time.

If you are choosing a language setting, read the guide to changing the dictation language and the guide to Mac dictation languages. Windows users can compare the Windows dictation languages list. If you want the general question first, see English voice typing for non-native speakers.