---
title: "What is Arabizi (Franco-Arabic)? Arabic in Latin letters"
description: "Arabizi is Arabic typed in Latin letters with digits such as 3 for ع. Why it is hard for NLP, how LLMs handle it, and why Latin-letter transcripts differ."
url: https://broadrouter.com/glossary/arabizi/
lang: en
updated: 2026-11-16
publisher: BroadRouter (a BroadNet site); API by BroadNet Inference, https://inference.broadnet.ai
---

# Arabizi

Updated 2026-11-16

Arabizi is Arabic written informally in Latin letters, using digits for Arabic sounds with no Latin equivalent, such as 3 for ع. It is common in chat and social media, is mostly dialectal, has no standard spelling, and is often mixed with English or French.

Also called: Franco-Arabic, Franco-Arab, Arabish, Arabic chat alphabet, Franco Arabic

Arabizi is how many Arabic speakers type on a phone: Arabic words, Latin keyboard, a few digits for the missing sounds.

## How it works

Darwish (2014) describes Arabizi, also called Arabish or Franco-Arab, as Arabic in transliterated Latin form that uses numerals for letters English lacks: “2” and “3” stand for the hamza sound and ع. It grew because older browsers and phones could not display Arabic script. It survives because people are used to it and type faster on an English keyboard.

Three properties make it hard for software:

-   **No standard spelling.** Shazal, Usman and Habash (2020) note that social-media Arabizi is dialectal Arabic, which has no standard orthography, and that the mapping to Arabic script is not one-to-one.
-   **Mixed with English or French.** A message like `w aletli Tuesday` (“and she told me Tuesday”) mixes both, which is [code-switching](https://broadrouter.com/glossary/code-switching/).
-   **Detection first.** Software must decide which words are Arabizi before converting them. Darwish reports 98.5% detection accuracy and 88.7% conversion accuracy with a classic pipeline; the 2020 sequence-to-sequence model reaches 80.6% word accuracy on a blind test set for the joint task.

For LLM APIs, Arabizi also behaves differently from Arabic script under the [tokenizer](https://broadrouter.com/glossary/tokenizer/), since it is Latin text spelled in ways the vocabulary rarely saw.

## Why it matters

If you run a chatbot, search or analytics in the Gulf, Egypt or the Levant, some of your users will type Arabizi. Keyword search, intent rules and sentiment tools built for Arabic script will miss those messages. A common fix is to ask an LLM to convert Arabizi to Arabic script, with the dialect named, before routing or indexing, and to keep the original text. The [Arabizi handling guide](https://broadrouter.com/blog/arabizi-llm-handling/) walks through prompts and reply-script choices.

## On BroadNet Inference

BroadNet has no Arabizi-specific feature; chat models receive it as ordinary text. BroadNet reports testing [Qwen3.8 27B](https://broadrouter.com/models/qwen3.8-27b/) in Arabic, with Gulf and Egyptian dialects understood, which is a starting point for Arabizi tests. See [Gulf and Egyptian dialects with LLMs](https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/).

Speech is a separate case. BroadNet’s docs warn that without `language`, Arabic audio can be transliterated into Latin letters. That output looks like Arabizi but comes from the request, not the speaker. Send `language=ar` to [BroadNet Transcribe Arabic](https://broadrouter.com/models/cohere-transcribe-arabic/).

## Questions

### Why does Arabizi use numbers?

Arabic has sounds and letters with no Latin equivalent, so writers borrow digits that look like the Arabic letter. Darwish (2014) gives 2 for the hamza sound and 3 for ع as examples.

### Can LLMs read Arabizi?

Often, but spelling varies so much that results are uneven, and conversion to Arabic script can differ from run to run. Test with real messages from your users and name the dialect in the prompt.

### My Arabic transcript came back in Latin letters. Is that Arabizi?

No. That is the speech model transliterating because the request did not say the audio was Arabic. On BroadNet, send language=ar.

## See also

-   [Code-switching](https://broadrouter.com/glossary/code-switching/) (term)
-   [Modern Standard Arabic (MSA)](https://broadrouter.com/glossary/msa/) (term)
-   [Tokenizer](https://broadrouter.com/glossary/tokenizer/) (term)
-   [Speech-to-text](https://broadrouter.com/glossary/speech-to-text/) (term)
-   [Dialectal Arabic](https://broadrouter.com/glossary/dialectal-arabic/) (term)
-   [Arabic speech-to-text API, step by step](https://broadrouter.com/guides/arabic-speech-to-text-api/) (guide)
-   [Arabizi and LLMs: handling Franco-Arabic with numbers like 3 and 7](https://broadrouter.com/blog/arabizi-llm-handling/) (article)
-   [Gulf, Egyptian and Levantine Arabic: how LLMs handle dialects](https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/) (article)
-   [Arabic prompt engineering: register, RTL, numerals and mixed text](https://broadrouter.com/blog/arabic-prompt-engineering/) (article)

Try it on your real workload. New accounts start with $5 of credit for Gemma and transcription. DeepSeek, GLM and Qwen3.8 need a top-up.

[Get started](https://inference.broadnet.ai/account/?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-arabizi&utm_content=page-cta) ·

**Sources**

-   [Darwish, Arabizi Detection and Conversion to Arabic (ANLP 2014, ACL Anthology) (checked 2026-11-16)](https://aclanthology.org/W14-3629/)
-   [Shazal, Usman and Habash, A Unified Model for Arabizi Detection and Transliteration (WANLP 2020, ACL Anthology) (checked 2026-11-16)](https://aclanthology.org/2020.wanlp-1.15/)
-   [BroadNet Inference API reference, transcription (language parameter) (checked 2026-11-16)](https://inference.broadnet.ai/docs?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-arabizi&utm_content=sources)
