BroadRouter

Dialectal Arabic

Updated

Dialectal Arabic is the everyday spoken Arabic of a region or city, such as Egyptian, Gulf, Levantine or Maghrebi Arabic. Dialects differ from Modern Standard Arabic in sounds, words and grammar, have no standard spelling, and are increasingly written online, which makes them harder for language models.

Also called: Arabic dialects, Arabic dialect, colloquial Arabic, Gulf Arabic, Egyptian Arabic, Levantine Arabic, Maghrebi Arabic, Khaleeji

A model can score well on Arabic benchmarks and still stumble on a voice note from Riyadh or a chat message from Casablanca. Those are dialects, and they are a different test from Modern Standard Arabic.

How it works

Bouamor et al. (2018), who built the MADAR corpus, describe Arabic as “a family of varieties”. MSA is the shared language of culture, media and education “from Morocco to Oman”, but it is not the native language of any speaker. Dialects are usually grouped regionally, as Egyptian, North African, Levantine, Gulf and Yemeni, or sub-regionally, as Tunisian, Lebanese, Kuwaiti and so on. They vary from MSA in sounds, vocabulary and morphology, and to a lesser extent from city to city.

Three differences matter most for software:

  • Sounds. MADAR’s example is the letter qaf (ق). In Tunisian Arabic it stays /q/, close to MSA; in Egyptian and Levantine Arabic it is a glottal stop; in Gulf Arabic it is /g/. Speech models must map all three back to the same written word.
  • Words. Everyday words differ. “I want” is أريد in MSA, عايز in Egyptian, بدي in Levantine, أبي or أبغى in Gulf Arabic and بغيت in Moroccan Arabic.
  • Spelling. MADAR notes that MSA has a standard orthography and the dialects do not. People spell by sound or by etymology, and sometimes in Latin letters as Arabizi. Speakers also mix in English or French, which is code-switching.

Dialect is increasingly written, too. MADAR calls it the emerging primary written language of informal communication online. NADI 2020 built a shared task from tweets covering 100 provinces in 21 Arab countries.

Why it matters

On BroadNet Inference

BroadNet lists BroadNet Transcribe Arabic for Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic. BroadNet reports a word error rate of 3.8 on FLEURS-ar read speech and 23.4 on its own Gulf, Egyptian and Levantine dialect set, both BroadNet measurements. The gap is the dialect effect. Send language=ar for all of them; BroadNet’s docs warn that Arabic can otherwise come back transliterated into Latin letters.

For text, BroadNet reports testing Qwen3.8 27B in Arabic with Gulf and Egyptian dialects understood. BroadNet does not publish a result for Levantine or Maghrebi text, so test those with your own messages. The Arabic LLM guide compares the models, and the Arabic speech-to-text guide has working code.

Common mistakes

  • Testing only on MSA news text, then shipping to users who write in dialect.
  • Averaging WER across dialects, which hides the weakest one.
  • Normalizing dialect spelling to MSA in the text you show users.

Questions

Which language code do I send for dialect audio?

Send ar. BroadNet's docs ask for language=ar on Arabic audio, and BroadNet lists Transcribe Arabic for Gulf, Egyptian, Levantine, Maghrebi and Modern Standard Arabic under that one model.

Should the model reply in dialect or MSA?

Decide per product and say so in the system prompt. Many assistants read dialect and answer in plain MSA or in the user's dialect; test which your users prefer.

How many Arabic dialects are there?

There is no fixed count. Dialects are grouped regionally (Egyptian, North African, Levantine, Gulf, Yemeni) or sub-regionally. MADAR alone covers 25 city dialects.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity