Speaking Arabic is not the same as speaking Khaleeji

A contact centre in Riyadh told us something we have thought about ever since. They had tested a voice assistant, and their verdict was not that it was wrong. It was that it knew what to do. It just did not sound Saudi.

That sentence is the whole problem in eight words.

Arabic on the box is not Arabic on the phone

Almost every voice product on the market supports Arabic. What they mean is Modern Standard Arabic, the Arabic of news broadcasts, official documents and school textbooks.

It is a real language and it is the right choice for a written notice. It is not how anyone in Riyadh, Jeddah, Kuwait City or Dubai asks about a delivery.

Ask a Saudi customer what they want and they will not say uridu. They will say abgha. A system trained on Modern Standard Arabic hears that as an unfamiliar word, or hears a word that sounds similar and means something else.

The same gap runs through ordinary speech. Numbers get said differently. Times get said differently. Greetings are completely different, and a greeting is the first thing a caller hears.

Sounding wrong is not a cosmetic problem

It is tempting to treat dialect as polish, something to fix in version two once the functionality works. That underestimates what happens on the call.

When an automated voice speaks in a register nobody uses on the phone, callers change how they talk. They slow down. They switch to English. They start repeating themselves before anything has gone wrong, because they have already decided the machine will not understand.

That behaviour is what breaks the call, and it starts in the first three seconds. By the time the system asks its first real question, the caller has already lost confidence in it.

Then they ask for a human, which is the outcome the whole system was bought to avoid.

Khaleeji is not one accent either

Treating the Gulf as a single dialect is the second mistake, made by people who have already avoided the first.

Saudi Arabia alone carries several distinct regional varieties. Kuwaiti, Emirati, Bahraini and Qatari speech each have their own vocabulary and rhythm. A contact centre in Riyadh serving customers across the GCC hears all of it in a single afternoon, and often within a single call, because Gulf speakers code-switch into English constantly. Half a sentence in Arabic, a product name in English, back to Arabic for the question.

A system that handles one dialect cleanly and falls over on the next is not solving the problem. It is relocating it.

Why this does not fix itself

The global models improve every year, and their Arabic improves with them. But they improve on the data that exists, and the overwhelming majority of recorded, transcribed Arabic is Modern Standard. Dialect speech is mostly spoken, rarely written down, and almost never labelled at scale.

That imbalance does not close on its own. It closes when somebody deliberately collects Gulf conversations, with consent, and trains on them.

This is the work Absher does. Every deployment, where the customer has opted in, makes the next call in that dialect better. It is slow, it is unglamorous, and it is the part that cannot be bought off a shelf.

What to test before you buy anything

If you are evaluating a voice system for the Gulf, do not test it in Modern Standard Arabic and do not test it yourself if you know what it expects to hear.

Ask three colleagues from different countries to call it and speak exactly as they would to a person. Let them switch to English mid-sentence. Let them use the words they actually use.

You will know within a minute whether the system was built for Arabic or built for your customers.

If you would like to run that test on Absher, get in touch.