Arabic and Emirati dialect
in voice AI.

For a business in the UAE, an AI receptionist that only handles English is half a receptionist. Many callers prefer Arabic, many switch between Arabic and English in the same sentence, and most speak a Gulf dialect rather than the Modern Standard Arabic of news broadcasts. Speech technology has improved enormously, but dialects remain the hardest part.

This article explains why, what published research measures, what vendors document, and how to test a voice agent with your own callers.

Why dialects are hard

Modern Standard Arabic (MSA) is the shared written and formal spoken language across the Arab world. Everyday speech is dialectal, and dialects differ from MSA and from each other in vocabulary, pronunciation and grammar. Emirati Arabic belongs to the Gulf, or Khaleeji, group, with local variation between emirates and communities.

Speech recognition models learn from transcribed audio. There is far more transcribed MSA, from news and formal speech, than transcribed Gulf conversation. Dialects are also written inconsistently, because there is no standard spelling for them, which makes both training and measuring accuracy harder. Add code-switching with English, names of people and places, and the narrow audio of a phone line, and the task gets tougher still.

What the research shows

Two recent open benchmarks give a sense of the gap, using open-source models.

• The Open Universal Arabic ASR Leaderboard (2024) evaluated open-source Arabic speech recognition across several test sets. On one multi-dialect set, the best model it reported scored a word error rate of about 19% on MSA but about 48% on Khaleeji speech. The authors note a significant decline on dialects such as Egyptian and Khaleeji, and caution against comparing dialects across different datasets.
• The Casablanca dataset (2024) covers eight Arabic dialects including Emirati. On its Emirati test set, widely used open models such as Whisper large-v3 produced word error rates above 50%. The authors report high error rates across every dialect they tested.

These are open models evaluated on research datasets, not the commercial systems you would deploy, and commercial systems trained on more dialect data may do much better. But the pattern is consistent: performance that looks excellent on MSA can fall sharply on Gulf speech.

What vendors document

A growing number of commercial services list UAE or Gulf Arabic separately, which is a good sign of dedicated work:

• Speech recognition: Deepgram's Nova-3 lists 17 Arabic variants, including ar-AE. Azure Speech lists 18 Arabic locales including ar-AE. Google's Chirp 3 lists ar-AE among its Arabic locales, all marked Preview. Speechmatics describes one Arabic model covering MSA and Gulf, Egyptian and Levantine speech, plus an Arabic–English bilingual option.
• Voices: Azure offers two UAE Arabic neural voices, and ElevenLabs lists Arabic (Saudi Arabia, UAE) on its multilingual models.
• Speech-to-speech models document Arabic less clearly. OpenAI's GPT-Live documentation does not list supported languages, and Google's Live API language table lists Egyptian Arabic but no Gulf locale. Test these directly.

A locale code tells you what has been packaged, not how well it works for your callers. Vendor accuracy figures are measured on the vendor's own test sets.

Understanding is only half of it

A good Arabic voice agent also has to sound right. Callers notice immediately if the agent replies in stiff formal Arabic, or in another country's dialect. At Clodia, the voice is instructed to use a natural Emirati register rather than formal newsreader Arabic or Egyptian or Levantine wording, to keep Arabic replies brief, and to use precise standard terms where safety or clarity requires it. Every phrase we add is reviewed in context, because dialect quality ultimately needs a native listener; a transcript can only verify wording.

How to test an Arabic voice agent

1. Use real phone calls, not browser demos, from mobiles in ordinary environments.
2. Recruit native speakers from the communities you serve, across ages and genders.
3. Script realistic calls: a booking, a complaint, a repair request with an address, a caller spelling a name, a caller reading a reference number.
4. Include code-switching, such as starting in English and finishing in Arabic, or mixing within a sentence.
5. Check names, numbers and places in the transcript, since these carry the facts.
6. Have a native speaker rate the replies for naturalness, politeness and correctness, not just the transcript accuracy.
7. Repeat after every change. A model update can improve one dialect and worsen another.

Practical tips for UAE businesses

• Greet in both languages, or in the language your callers most often use, and follow the caller's lead.
• Confirm important details back to the caller, especially numbers and names.
• Give the agent a short list of your own names, products and locations, and test them.
• Make it easy to reach a person, and review Arabic calls more often in the first weeks.

Read this article's Arabic companion: الذكاء الاصطناعي الصوتي واللهجة الإماراتية, or our blog post on bilingual voice AI.