Northline AI

Gulf Arabic, code-switching and voice AI: why dialect matters

Most callers in the UAE do not speak textbook Arabic, and many switch to English mid-sentence. What speech-recognition research says and what to test before you buy.

Northline AI ResearchEditorial team, reviewed by Abdelfattah Alsaeed
·4 min read
37.5%best word error rate on dialectal Arabic in the MGB-3 challenge

Key takeaways

  • Callers in the UAE mostly speak Gulf (Khaleeji) and other spoken dialects, not the Modern Standard Arabic most AI is trained to read.
  • In the Arabic MGB-3 challenge, the best system still had a 37.5% word error rate on real-world dialectal Arabic, compared with 13.2% on broadcast news.
  • Code-switching between Arabic and English mid-sentence is normal in the UAE, where almost everyone is online and mobile.
  • Test any Arabic voice agent with real dialect, mixed language and local names before you buy.

"We support Arabic" is on almost every AI vendor’s website. But which Arabic? A customer in Al Ain, a resident from Cairo, a Lebanese manager in Dubai Marina and a Saudi visitor will all speak to your business very differently. And many of them will switch to English halfway through a sentence. If your voice agent only handles textbook Arabic, it does not really handle Arabic customers.

Arabic is many spoken languages

Arabic has a written standard, Modern Standard Arabic (MSA), used in news, official documents and formal writing. But people do not speak MSA at home or on the phone. They speak regional dialects: Gulf (Khaleeji), Egyptian, Levantine, Iraqi, Maghrebi and more. The differences are not just accent. Everyday words change completely.

MeaningMSAGulf (Khaleeji)EgyptianLevantine
I wantأريدأبي / أبغىعايزبدي
Nowالآنالحيندلوقتيهلق
What?ماذا؟شو؟ / وش؟إيه؟شو؟
Good / OKجيدزينكويسمنيح

A system trained mostly on MSA text will struggle with "أبي موعد الحين" even though any Emirati receptionist understands it instantly.

What the research shows about dialect recognition

Speech recognition for dialectal Arabic has been a known hard problem in academic research. The Arabic Multi-Genre Broadcast challenge (MGB-3), organised by researchers from the Qatar Computing Research Institute and the University of Edinburgh, tested systems on real Egyptian Arabic speech from YouTube.[1]

37.5%
best average word error rate on real-world dialectal Arabic in MGB-3Ali, Vogel & Renals (QCRI / University of Edinburgh), 2017
  • The best system on the dialect test had an average word error rate of 37.5%.
  • On the broadcast news test set from the earlier MGB-2 challenge, which is mostly MSA, the best result was 13.2%.
  • For identifying which of five dialects was spoken, the best system reached about 75–80% accuracy.

Technology has improved a great deal since 2017, especially with large multilingual speech models. But the gap the challenge exposed has not disappeared: the further speech moves from formal, scripted Arabic towards casual, regional, noisy phone audio, the harder it gets. That is exactly what business phone calls are.

Code-switching is normal in the UAE

The UAE is one of the most connected countries on earth. DataReportal’s Digital 2025 report puts internet use at 99% of the population, with mobile connections equal to 195% of the population.[2] Its workforce is international, and business happens in a mix of Arabic and English, sometimes in the same sentence:

أبي أحجز appointment يوم Thursday بعد الساعة five إذا ممكن

A voice agent has to understand that sentence as one request, not two languages. It also needs to know which language to reply in. A good default: answer in the language the customer started with, and follow them if they switch.

Local knowledge matters as much as language

Understanding Arabic is not enough if the agent does not know local names and conventions:

  • Places: Al Barsha, Mirdif, Al Nahda, Khalifa City, Muwaileh.
  • Names: Arabic, South Asian, Filipino, Russian and European names, often spelled several ways.
  • Formats: UAE phone numbers, Emirates ID references, AED prices, the Monday-to-Friday working week.
  • Courtesy: greetings like "السلام عليكم" and "الله يعافيك" are part of good service, not small talk.

How to test an Arabic voice agent before you buy

  1. Call it in your real dialect, not formal Arabic. Use everyday words like "أبي" and "الحين".
  2. Mix languages the way your customers do.
  3. Say local names: your area, a common surname, a nearby landmark.
  4. Give a phone number and a date in Arabic. Check it is captured correctly.
  5. Call from a noisy place, like a car or a mall. Real calls are not studio recordings.
  6. Listen to the reply. Does it sound Gulf, or like a news anchor?

Hear the agent in Khaleeji Arabic on your own number.

See how it works

The bottom line

Arabic support is not a checkbox. For a business in the UAE, the real question is whether a caller from down the road can speak naturally, in their own dialect, mixing English where they like, and be understood the first time. Test for that, and you will find out quickly which systems were built for this region.

Frequently asked questions

Yes, if it has been trained and tested for it. Many systems handle Modern Standard Arabic far better than spoken dialects, so test any voice agent with real Khaleeji phrases, local names and mixed Arabic-English before you buy.

Sources

Every statistic in this article comes from the studies below.

  1. 1.Speech Recognition Challenge in the Wild: Arabic MGB-3 Ali, Vogel & Renals (QCRI / University of Edinburgh), 2017 Peer-reviewed
  2. 2.Digital 2025: The United Arab Emirates DataReportal (Kepios, We Are Social, Meltwater), 2025

Researched by

Northline AI Research

Editorial team, reviewed by Abdelfattah Alsaeed

The Northline AI research team reads the studies so UAE business owners do not have to. Every statistic links to its original source.

Keep reading

Never lose another customer to an unanswered call

Launch your AI receptionist in minutes in our self-serve cloud console, or forward your after-hours calls and judge the results for 14 days.