The Arabic gap: why most voice AI fails in MENA markets

Here is why I know this model will make a difference in Arabic-speaking countries.

 

A language that does not behave like other languages

 

Arabic is not a single language in any practical sense. Gulf, Egyptian, Levantine and Maghrebi dialects carry distinct vocabulary, phonology and rhythm.

 

Different people will pronounce the same word differently. The variation is comparable to a thick Boston accent and a deep Mississippi drawl.

 

I‘m Egyptian. My accent carries differently to a Gulf speaker’s, and both of us carry differently to someone from Tunis or Beirut. A model trained primarily on broadcast Modern Standard Arabic will struggle with many real conversations.

 

We made sure we had examples from different dialects and dialect-specific training data precisely because enterprises need products and services that travel across borders.

 

Code-switching: the wall no one sees coming

 

In most Arabic-speaking countries, people often switch between English and their dialect, particularly between formal speech and conversational speech. It is the default register for educated professionals across the region, such as a doctor naming a drug in English or a finance officer mixing both in a single sentence.

 

The Gulf also has one of the highest concentrations of expat communities in the world, where Arabic and English are the working languages. If someone switches mid-sentence, running two models in parallel does not fix it. If you run two models, one is going to think this is English and one‘s going to think it’s Arabic, and to combine them they won’t know which one to listen to.

 

For optimum accuracy and speed, you need a single model that knows both languages simultaneously and can differentiate between the words as they are being spoken.

 

Specialised models: when errors stop being typos

 

In regulated sectors, a transcription error is a potential liability. Nowhere more so than medical transcription, where a weak model that drops a drug name puts that error straight into the patient record, creating more work for already stretched doctors.

 

A dedicated domain model does not just reduce those errors. It allows us to achieve even higher accuracy in real settings, where vocabulary is specialised and code-switching is constant.

 

As voice agents move into clinical and regulated environments, the tolerance for that gap moves toward zero. That is why we built a dedicated Arabic-English medical model trained on clinical vocabulary and real code-switching.

 

You can see our new medical model tested in the real world by a retired Egyptian anaesthetist (also my father!), below.

 

 

Where the model runs matters as much as how it performs

 

Enterprises across the region are asking whose infrastructure their sensitive data runs on. Saudi Arabia’s Personal Data Protection Law came into full enforcement in September 2024, and the UAE’s Federal Data Protection Law has been in force since January 2022.

 

Both have concrete implications for where voice data is processed and stored, and cross-border processing triggers obligations many vendors cannot meet.

 

Beyond compliance, enterprises across the region are asking whose infrastructure their sensitive data runs on. The current geopolitical climate has made that question harder to ignore, and deployments here reflect it: on-prem and on-device, built around sovereignty from the outset.

 

The infrastructure is arriving. The linguistics need to keep up.

 

All these things are tangible improvements that will really make a difference: in hospital systems, in government service fronts, in the contact centres handling tens of millions of interactions a year across the region.

 

But the models have to work for the people using them. The gap is measurable, but it is closing.

---

Read more about Arabic transcription on speechmatics.com

---

By Yahia Abaza, Senior Product Manager, Speechmatics