When people say an AI model “speaks Arabic”, they usually mean it handles Modern Standard Arabic. Dialects are a separate test.
How it works
Arabic is a family of varieties. The MADAR paper (Bouamor et al., 2018) puts it plainly: MSA is “the shared language of culture, media, and education from Morocco to Oman”, yet it “is not the native language of any speaker of Arabic”. Dialects are grouped regionally as Egyptian, North African, Levantine, Gulf and Yemeni, and sub-regionally as Tunisian, Lebanese, Kuwaiti, Qatari and so on. MADAR itself covers 25 city dialects alongside MSA.
In practice the two mix. In unscripted settings where MSA would be expected, such as TV talk shows, speakers switch back and forth between MSA and their dialect. That is code-switching, and the 2014 code-switching shared task found MSA versus dialect the hardest pair to tell apart at the word level, harder than Spanish-English or Nepali-English.
| MSA | Dialects | |
|---|---|---|
| Where | news, books, school, contracts, formal speech | home, street, phone calls, voice notes, chat |
| Spelling | standard | no fixed spelling; often typed as Arabizi |
| Training data | large written corpora | much less, unevenly spread by region |
Why it matters
- LLM quality. A model can write good MSA and still miss Gulf or Maghrebi slang. Test with the dialect your users speak.
- Speech-to-text accuracy. Many test sets are read or broadcast speech, while your audio may be dialect voice notes. Ask for word error rate by variety, not one Arabic number, and test on your own recordings.
- Token cost. In our tokenizer dataset, on Qwen3.8 27B Modern Standard Arabic needed 1.31× the tokens of the same sentences in English (2,009 FLORES-200 sentences) and Gulf Arabic 1.21× (a small set of 49 Tatoeba sentences, so a rough figure). Use the Arabic token counter on your own text.
On BroadNet Inference
BroadNet reports testing Qwen3.8 27B for fluent Arabic writing, with Gulf and Egyptian dialects understood. BroadNet Transcribe Arabic is listed for Arabic including dialects. For both MSA and dialect audio, send language=ar; BroadNet’s docs warn that without it Arabic can come back in Latin letters. The dialect speech recognition article explains how to test on your own recordings.