What is the best Arabic speech to text API?

No single API is best for Arabic. Google Cloud, Microsoft Azure, Amazon Transcribe, Deepgram and OpenAI's Whisper all accept Arabic, but they list different varieties, and the accuracy figures published so far come from different tests. The right choice depends on the variety you speak, where your audio is processed and how you will use the text. Any page that names one winner has not run the same test on the same speech.

Each provider lists different Arabic varieties

• Google Cloud Speech-to-Text lists Arabic (Egypt) with the code ar-EG for its chirp_3 model ¹.
• Microsoft Azure Speech lists separate Arabic locales for Egypt (ar-EG) and Morocco (ar-MA) ².
• Amazon Transcribe lists Arabic (Gulf) as ar-AE and Arabic (Modern Standard) as ar-SA ³. Its table also shows which request modes each entry supports, so check the entry you need.
• Deepgram gives Arabic the base code ar and separate regional codes, including ar-EG ⁴.
• OpenAI's Whisper has Arabic in its language list, under the code ar ⁵.

The published accuracy figures come from different tests

Accuracy is measured on one test set at a time. On the Fleurs test, Whisper large-v2 had a word error rate of 16.0 percent for Arabic ⁶. A comparison run by Mistral, on one read-speech test set, gave GPT-4o mini Transcribe a word error rate of 14.02 percent for Arabic ⁷. The two numbers are not directly comparable. One is for a 2022 model on a public benchmark, and the other is a vendor's own test on read speech.

Neither figure measures dialect speech. On traditional Emirati Arabic, the Whisper models of a 2025 study reached 86 to 93 percent word error rate without training on the dialect ⁸. None of the APIs above was tested in that study, so the figure describes the Whisper models only.

Choose by variety, then by data terms

• Egyptian Arabic: the lists above show ar-EG on Google, Azure and Deepgram.
• Gulf Arabic: the Amazon list has ar-AE.
• Modern Standard Arabic: Amazon lists ar-SA, and Whisper's language list includes Arabic.

Check each provider's data terms before you send recordings, because they differ in whether they keep audio. The question on whether speech to text APIs store audio covers the settings to look for. If the audio must stay on your own computer, self-hosted Whisper is the route, and the question on running Whisper offline explains what that takes.

Run a test before you commit

1. Record 10 minutes of your own speech in the variety you use, with the microphone you will use, including the English words you normally mix in.
2. Send the same file to two or three providers, each with the closest Arabic code.
3. Correct one reference transcript by hand, then compare each provider's word error rate, using the method for calculating it.
4. Read the errors that matter most to you, such as numbers, names and English terms.

Choose the lowest error rate that also fits your data terms and budget. A single test of 10 minutes is a starting point, not proof, so repeat it with a second recording on another day.

It depends on what the text is for

• Dictating into a document yourself: the built-in options on your phone or computer may be enough, and they need no integration. Our Arabic guide lists which Arabic entries each one offers.
• Building a product: check whether the provider supports batch or streaming for your locale, then test latency with your own audio as well as accuracy.
• Mostly English with some Arabic: an API that treats Arabic as a secondary language may be weaker on the Arabic parts, so test the mix you actually speak.