Arabic Speech to Text and Right-to-Left Dictation

You can dictate Arabic on a Mac, an iPhone, an iPad and in Microsoft Word, and the built-in options cover a handful of country variants. Windows 11 voice typing does not offer Arabic at all. The bigger problem is accuracy: speech models make many more errors on everyday dialects than on Modern Standard Arabic ¹. Right-to-left script adds a second problem, because numbers, English words and punctuation can appear in an order you did not speak.

Yap, which we make, is an on-device dictation app for Mac and Windows. Check our download page for what it supports today, and treat the lists below as the built-in options on the platforms named.

Arabic dictation is built into Apple devices and Word, but only for a few country variants

Apple's dictation lists name two Arabic variants on the Mac: Arabic (Saudi Arabia) and Arabic (United Arab Emirates) ². The iPhone and iPad list has the same two ³. Microsoft Word lists three Arabic variants, Bahrain, Egypt and Saudi Arabia, in its Preview group ⁴. The same page warns that Preview languages may have lower accuracy or limited punctuation support ⁵. So Word supports Arabic, but Microsoft warns that these Preview languages may be less accurate.

Windows 11 voice typing lists 46 languages and locales, from Bulgarian to Vietnamese ⁶. Arabic is not on that list, so Windows 11 voice typing does not dictate Arabic. Our Windows voice typing guide covers what the feature does offer.

The table below puts the four languages of this guide side by side on each platform, using the lists we read on 2026-10-17.

Two cautions apply to that table. First, the iPhone and iPad on-device list names Arabic with no country, so the page does not say which Arabic variant runs without a connection ⁷. Second, the Azure row is a developer service, not a consumer dictation app. Its locale table includes Hebrew (Israel) ⁸, Persian (Iran) ⁹ and Urdu (India) ¹⁰, and the consumer lists above include neither Persian nor Urdu.

Whisper, the open speech model from OpenAI, includes Arabic in its language table ¹¹. If you run Whisper yourself, you are not limited to a platform's list. Our guide to Whisper's supported languages explains how its accuracy varies by language.

Dialects are the biggest accuracy problem in Arabic recognition

The main gap in Arabic dictation is dialect, not whether Arabic is offered at all. A 2025 preprint on Arabic recognition without training on each variety reports that error rates rise sharply on dialectal and accented speech, and it attributes the problem to linguistic mismatch and scarce labeled data ¹. Platform lists name country variants, not dialects, so a speaker of a dialect has to choose the nearest entry. Modern Standard Arabic is the official language of spoken and written Arabic media ¹², while colloquial Arabic covers the regional spoken varieties.

The gap is not new. On the 1996/97 NIST CallHome task for Egyptian Colloquial Arabic, the best word error rate reached 61 percent ¹³. That system is far older than today's models, so the number shows the size of the problem rather than the error rate you will see now.

Dialect test sets also differ from one another. The 2017 MGB-3 challenge emphasized dialectal Arabic, drawing its data from Egyptian YouTube videos ¹⁴. A test built from news can say little about comedy, cooking or a family conversation, which is why one Arabic accuracy figure rarely tells you how a dictation tool will behave for you.

The Whisper paper shows the same spread on two test sets. Whisper large-v2 had a word error rate of 16.0 percent for Arabic on the Fleurs test ¹⁵, and 53.8 percent for Arabic on the Common Voice 9 test ¹⁶. The two sets have different speakers and recording conditions, so the gap should not be read as a dialect effect alone. It does show that the same model can look very different depending on the test.

Mixing is a separate problem. Switching into English inside a sentence is called code-switching. The authors of a 2023 study say that multilingual models such as Whisper often struggle with that setting ¹⁷. When you dictate mixed Arabic and English, proofread the English words and the punctuation around them first.

Our question on which Arabic dialects speech to text understands covers the practical side in more detail.

Pick the country variant closest to how you speak

Choose the Arabic entry closest to your own speech, even if it is not your country. The platform lists give you only a few choices, and the locale table gives you more. Azure Speech lists separate Arabic locales for many countries, including Egypt (ar-EG) and Morocco (ar-MA) ¹⁸. A locale tells the system which language and region to use, but the table does not show accuracy for each country, so it cannot tell you how well a given dialect will be recognized.

On a Mac running macOS 27, the language setting for Dictation sits in Keyboard settings. Apple's steps are:

1. Open Keyboard settings.
2. Select Dictation.
3. Click Edit next to Languages, then select the language and region you want ¹⁹.

Apple's own troubleshooting checklist for Dictation also tells you to choose the correct language and region when Dictation does not work as expected ²⁰. If the entry you choose does not match your speech, check a test paragraph before you rely on it.

For a quick test, dictate two or three minutes of your own speech in the variant you use every day, then count the errors. That tells you more than any table. Our guide to dictation languages on Mac lists the variants Apple offers in each language, and the guide to change dictation language covers switching on other devices.

Right-to-left text with numbers or English words can display out of order

Arabic and Hebrew are written right to left, while digits and Latin-script words are written left to right, so mixed text runs in both directions. Each run keeps its own internal order, and the boundaries between runs are where things go wrong. Our question on why dictated Arabic or Hebrew looks backwards gives the short answer.

The standard sets the rules for those boundaries. For a Western digit that follows an Arabic letter, the rule searches backward for the nearest strong character, and if that is an Arabic letter, it changes the digit's type to an Arabic number ²¹. Punctuation and spaces, which the standard calls neutral characters, generally take the direction of the text around them. When the two sides conflict, they take the embedding direction ²².

In practice, a comma or a closing parenthesis that sits between an Arabic word and an English word or a number can end up on the side you did not expect. What moves is mostly the punctuation and spaces at the boundaries. How an app applies these rules depends on its text engine, so the same sentence can look different in two apps.

When you proofread dictated Arabic that mixes in numbers or English:

1. Read the punctuation next to every number and every English word first, since those are the places where the order changes.
2. Check that a parenthesis or quotation mark closes the expression you meant it to close.
3. For a long document, proofread in the app you will publish from, because the display may differ between a note and a word processor.

Our guide to punctuation commands in other languages covers how spoken punctuation works in the languages that use these rules.

Dictated Arabic usually leaves out short-vowel diacritics

The marks that show short vowels are called diacritics. A 2023 Interspeech paper reports that most people write and type Arabic text without them ²³.

That is why dictation normally gives you letters without short vowels. You should not expect a speech tool to add the marks in a Qur'an text, a children's reader or a poem. We did not find a dictation product in the lists above that adds short-vowel marks on its own.

Diacritics can be added to a model's output, but where you add them matters. The 2023 study compared adding diacritics inside the speech model with adding them to the text afterward. It found that diacritizing inside the model significantly outperforms diacritizing the text afterward, particularly when the model is fine-tuned on manually diacritized transcripts ²⁴. That result comes from a research setup, not from a consumer app, so it tells you what works rather than what you can buy.

If you run Whisper yourself, you can test whether a sample prompt written with short-vowel marks changes its output on your own audio. We did not find a published result for Arabic, so treat that as a test to run, not a method that is known to work.

Hebrew has the same pattern. A 2025 preprint on restoring Hebrew diacritics says they are typically omitted in modern Hebrew writing ²⁵.

Hebrew, Persian and Urdu have their own support gaps

Hebrew has the most consumer support of the three. The Mac dictation list includes Hebrew (Israel) ²⁶, and Word lists it in the same Preview group as Arabic ²⁷. Hebrew is still behind Arabic on the test set Whisper used. Whisper large-v2 had a word error rate of 27.1 percent for Hebrew on Fleurs ²⁸, against 16.0 percent for Arabic ¹⁵.

Persian has the weakest consumer support. None of the Apple, Windows or Word lists we read includes Persian. The developer service does: Azure's locale table lists Persian (Iran) as fa-IR ⁹. On the Fleurs test, Whisper large-v2 had a word error rate of 32.9 percent for Persian ²⁹. Its English figure on the same test was 4.2 percent ³⁰, so the Persian rate is about 7.8 times English's ³¹.

Urdu sits in the middle. Windows voice typing does not list it, and Azure's locale table includes Urdu (India) as ur-IN ¹⁰. Whisper large-v2 had a word error rate of 22.6 percent for Urdu on Fleurs ³². A 2024 benchmark of Urdu models found that Seamless-large was best on read speech, while Whisper-large was best on conversational speech ³³. So the right model depends on whether you dictate a letter or talk through a meeting.

The chart compares the Whisper large-v2 Fleurs error rates for the five languages discussed above. Fleurs is a benchmark, so these bars are not a forecast for your own dictation.

Arabic's figure on the same test is about 3.8 times English's ³⁴. The gap is smaller than Persian's, but it is still there.

Typing is the fallback when no dictation option covers your variant

Dictation is not the only way to get Arabic text. If the built-in options do not cover your variant, typing is the fallback, and typing takes practice. In a 2024 study of Arabic learners, a mnemonic method raised typing speed from 12 to 25 words per minute and accuracy from 55 to 85 percent ³⁵. That is one study by one author, with 30 students ³⁶, so treat it as early evidence. Our guide to touch typing explains what touch typing is and how to learn it, and the question on learning touch typing in Arabic covers the Arabic keyboard specifically.

For the other questions this guide raises:

• Does speech to text understand Arabic dialects?
• What is the best Arabic speech to text API?
• Why does dictated Arabic or Hebrew text look backwards?
• How do you learn touch typing in Arabic?

For the wider picture, start with our overview of dictation languages, accents and voice translation. Related guides in this section are dictation languages on each device, speech recognition and accents and multilingual dictation. For the wider concepts, read the glossary entries on right-to-left script, locale and code-switching, which matters when you mix Arabic and English in one sentence.

If you want to dictate in your browser without installing anything, free dictation online works in the browser, and Mac dictation language help covers changing the language on a Mac.