What languages does Whisper support?

Whisper accepts 98 languages, but "supports" means something narrower than it sounds. The model has a language code for each of them, and it will try to transcribe speech in any of them. How well it does varies a great deal. Hebrew and Persian are in the list but produce error rates several times higher than English on the standard benchmark. Kurdish has no code in Whisper's language table, and Quran recitation has only small research studies. Test the language you need on your own audio before you rely on it.

Whisper accepts 98 languages, and OpenAI warns that accuracy varies

OpenAI's speech-to-text guide says Whisper supports 98 languages, and that accuracy varies by language ¹. The count is the number of languages the model accepts, not a measure of how well it handles each one. OpenAI's Whisper README says the same thing in other words: its performance varies widely depending on the language ².

The language table in the open-source code is the list of codes the model accepts, such as he for Hebrew, fa for Persian and id for Indonesian. You can read the table in the tokenizer source file, where Indonesian is written as "indonesian". The table is a list of codes, not a list of accuracy results.

Hebrew and Persian work, but with error rates several times higher than English

The clearest way to see the spread is the word error rate on the FLEURS benchmark, which reads sentences aloud in many languages. For the large-v2 model, Whisper's rate was 27.1 percent for Hebrew ³ and 32.9 percent for Persian ⁴. A word error rate of 27 percent means roughly one word in four is wrong, which is usable for skimming and not for a legal or medical record.

The same model's rate for English on that benchmark was 4.2 percent ⁵, so the gap is large. The word error rate counts substituted, deleted and inserted words against a reference transcript, so a figure for read sentences in a quiet room will understate the errors in everyday speech.

Hebrew also has a writing issue that the benchmark does not measure. A 2022 study notes that most Hebrew vowels are encoded as diacritics, so the pronunciation of a word is left underspecified in ordinary text ⁶. A transcript that omits those marks can still be correct, but a reference that includes them will score it as wrong. Decide before you test whether the marks matter for your use.

Arabic is in the table, but the Quran is studied only in small tests

Arabic has its own code, ar, in Whisper's language table in the source file. That single code covers several dialects, and the table does not separate them. An app built on Whisper may offer only some of them, so check which dialect the app transcribes.

Quran recitation is the case most people search for. A 2025 study applied Whisper to 200 verses from Juz 30, and reported an average character error rate of 0.142, which the authors describe as about 85 percent transcription accuracy ⁷. Character error rate counts letters rather than words, so the figure cannot be compared directly with word error rates from other studies. The abstract reports only that measure, so it says nothing about tajweed or the pronunciation rules that reciters follow. Treat the result as a single small study on one part of the Quran.

Kurdish has no code in the table, and Kurmanji results need fine-tuning

Kurdish is the language people most often ask about, and it is the one that is hardest to answer from the standard model. The language table in the open-source code has no Kurdish code, so you cannot select Kurdish as a language in the standard model. Speakers of Kurdish usually have to rely on an app that adds its own support, or on a model that someone has trained for the language.

Research shows what that training can achieve. A 2024 study fine-tuned Whisper version 3 on about 68 hours of Northern Kurdish (Kurmanji) speech, and reached a word error rate of 10.5 percent on a specialized test set ⁸. That is a result for one corpus and one test set, not a promise for everyday Kurmanji speech, and it describes a fine-tuned model rather than the standard one. Sorani, the other major variety, is not covered by a Whisper result in the sources we checked.

Indonesian and other codes are listed, so test before you rely on them

Indonesian is in the table under the code id, and Bengali and Urdu are also listed. Being in the table shows only that the model accepts the language. It does not show how accurate the model is for that language, and the benchmark results for Hebrew and Persian show that accuracy can be far from English.

For most languages, the honest answer is to test. Record two or three minutes of the speech you will use, correct a transcript by hand, and count the errors that matter to you, such as names and numbers. The steps are in our guide to speech recognition accuracy, and the guide to Whisper AI explains where the model's limits come from.

What to do if your language is missing or weak

1. Check the language code in the app you use. An app built on Whisper may offer fewer languages than the model accepts, so the language list in the app is what you can actually choose.
2. Test on your own audio. Use the same microphone, room and speaking style you will use later, and compare the results across two or three candidates.
3. Plan for correction. For Hebrew, Persian or any language with a high error rate, budget time to edit the output, especially for names and terms.
4. For Kurdish, look for an app or model built for the language. The standard Whisper model's table has no Kurdish code, so do not expect a standard setting to work.

For a wider comparison of how well other speech systems handle these languages, see the guide to speech recognition models, and which dictation app supports the most languages. Our question on whether Whisper's language limits apply to other models covers the general pattern.