How do you transcribe Arabic or Hebrew audio to text?

Set the spoken language to Arabic or Hebrew, transcribe the recording with a tool that supports that language, and then read the result against the audio. Expect to correct a real share of the words. In a published test on read sentences, Whisper large-v2 had a word error rate of 16.0 percent in Arabic ¹ and 27.1 percent in Hebrew ². Those are best-case figures. A conversation, a voice note recorded in a noisy street or a regional dialect can do worse. This page covers the steps, the dialect and vowel issues that cause most errors, and the checks to run before you use the text. For the wider picture, start with the guide to transcribing audio in other languages.

Arabic's error rate is about 3.8 times English's on the same test

A word error rate counts the words that are substituted, deleted or inserted, compared with a correct reference. A rate of 16.0 percent means that roughly one word in six is wrong. On this test, Whisper large-v2's Arabic error rate was about 3.8 times its English rate ³, and Hebrew's figure is higher than Arabic's.

Newer systems change the order. In a 2025 comparison on the same read-speech test, Gemini 2.5 Flash had an Arabic error rate about 64 percent higher than Whisper large-v3 ⁴. In Hindi the ranking reverses: Whisper large-v3's error rate was about 4.3 times that of Gemini 2.5 Flash ⁵. The lesson is to test the tool you plan to use on Arabic, not to pick one by reputation.

These figures come from read speech. Read speech is easier than most real recordings, so treat them as a floor for the error you should expect.

Arabic accuracy depends on the dialect

Arabic has a formal written and broadcast form, Modern Standard Arabic, and a set of regional spoken dialects that differ from it and from each other. Many recordings are in a dialect, and a model trained mostly on formal speech handles dialect poorly. In a 2025 study of traditional Emirati Arabic, Whisper Medium had a word error rate of 86.10 percent and Whisper Small of 93.06 percent when used without any training on that dialect ⁶. Those error rates are high enough that a dialect recording needs a plan before you start.

• If the tool lets you choose a regional setting, choose the one closest to the speaker.
• If the speaker switches between dialect and formal Arabic, check the switch points, where errors cluster.
• If the recording is important, have a native speaker of that dialect review the transcript.

Arabic and Hebrew are usually written without their vowels

Most people write and type Arabic without diacritics, the short vowel marks that tell you how a word is pronounced ⁷. Modern Hebrew is written the same way, and its diacritics, including the vowel marks, are typically omitted ⁸. The result is text with the consonants but without the vowel marks, and the recognizer has to guess the vowels from context.

That guess is a common source of errors. Two different words can look identical on the page, and a transcript may pick the wrong one. Check names, places and verbs against the audio rather than trusting the spelling. For the term, see diacritics.

Choose the regional setting for Hebrew and Arabic

Setting the language is the first step, and the regional setting is the second. Apple's Mac Dictation feature availability list includes Hebrew (Israel) ⁹, so choose that entry when the tool offers it. Some cloud services split Arabic into separate regional codes, such as Egypt (ar-EG) and Morocco (ar-MA) in Microsoft's Azure Speech table ¹⁰. Pick the code that matches the speaker, because a mismatch can change the vocabulary and spelling the model expects.

Do not rely on automatic detection for a short clip or a mixed recording. A wrong guess about the language can affect the whole transcript, so set the language yourself whenever you know it.

Right-to-left text needs its own check

Arabic and Hebrew are written right to left, but numbers, English names and brand names are written left to right. The Unicode standard treats text that mixes these directions as bidirectional, a mixture of right-to-left and left-to-right text ¹¹. When you paste a transcript into another app, these pieces can end up in the wrong order.

Read the first and last word of each line, where direction problems are easiest to see. Check numbers, dates and Latin-script names against the audio separately. For the wider issue, see right-to-left script.

When the recording mixes languages or matters legally

If the speaker mixes Arabic or Hebrew with English, expect more errors at each switch. Mixed-language speech, called code-switching, is harder for speech models ¹², so pick the dominant language and read the switch points closely.

For a medical instruction, a contract, a court statement or a recording that will be published, a human review is the safe choice. An AI draft can save typing time, but the final check should be by someone who speaks the language and dialect.

What to do, step by step

1. Choose the language and the regional setting before you transcribe, and use the closest match to the speaker.
2. Transcribe the first minute or two and read it. If the script or the words look wrong, change the setting and run it again.
3. Test on your own audio. Choose three to five recordings that include the hardest ones, correct one by hand to make a reference, and count the errors to get your own word error rate.
4. Proofread the whole transcript against the audio at several points, paying most attention to names, numbers and vowels.
5. Decide what an error would cost. If a wrong word could change the meaning, plan a human review.

For a file you can upload in a browser, our free audio to text tool works from the file itself. Yap, which we make, is an on-device dictation app for Mac and Windows, so if you want to dictate live, check the download page for the languages it supports.

Other language questions in this guide: how to transcribe Hindi audio, Tamil, Urdu or Bengali audio, Spanish, German or Dutch audio, and translating a voice recording to English. For the wider method, see how to transcribe audio to text and the Arabic speech to text guide.