Arabic Voice Agents for Gulf Hospitals: Dialects, Mixed English Speech and Number Handling

This post is for marketing and patient-access leaders at hospitals and clinic groups in the UAE and wider Gulf who are considering an AI voice agent for their call line, and for the technical teams who will configure it. It answers three practical questions. Which Arabic should the agent understand and speak? How should it cope with callers who switch to English mid-sentence? And how should it read back phone numbers, dates and times so a booking is actually correct?

Arabic is not one spoken language in the way a hospital's English front desk script is one language. What is written in a newspaper, what is said at a reception desk in Dubai and what is said in Beirut or Cairo differ in vocabulary, pronunciation and grammar. A voice agent trained or configured for the wrong variety will misunderstand callers or, almost as bad, sound foreign to them.

This is an explainer built on how speech systems are specified, not a set of benchmark results. We recommend a test-call protocol and leave the measuring to you, because performance varies by vendor and changes with each model release.

The varieties of Arabic a hospital line meets

Think in four layers.

• Modern Standard Arabic (MSA). The formal written and broadcast register, shared across the Arab world. People understand it everywhere, but few use it for a phone call to a hospital. It is the right language for written content, official notices and formal announcements, and a reasonable fallback register for an agent's replies.
• Gulf Arabic (Khaleeji). The everyday speech of Emirati, Saudi, Kuwaiti, Qatari, Bahraini and Omani callers, with differences between and within those groups. Expect pronunciation features and vocabulary that MSA-trained systems may not handle.
• Levantine Arabic. Spoken by many residents and healthcare staff from Lebanon, Syria, Jordan and Palestine, and common in the Gulf's expatriate population.
• Egyptian Arabic. The most widely understood dialect through film and television, and the native speech of a large expatriate group in the Gulf.

Gulf hospitals also serve large South Asian and Filipino populations who speak English, Hindi, Urdu, Malayalam, Tagalog and others. That is outside this post but relevant to your language menu.

For a typical Gulf hospital, the practical conclusion is that the recogniser has to accept several dialects as input, while the agent's spoken replies should be in a register the caller finds natural and professional. Many teams choose a light, clear Gulf-leaning or neutral register for output rather than strict MSA, which can sound stiff, or a heavy dialect, which can sound inconsistent. This is a brand decision as much as a technical one and should be tested with native speakers.

What speech recognisers actually offer

Recognition services typically list Arabic by locale. Google's supported-languages page for Speech-to-Text lists many Arabic locales, including ar-AE for the UAE, ar-SA for Saudi Arabia, ar-KW for Kuwait, ar-QA for Qatar, ar-BH for Bahrain, ar-OM for Oman, ar-EG for Egypt, ar-JO for Jordan and ar-LB for Lebanon. A locale code does not by itself tell you how well the model understands that dialect, and feature support varies by model and region, so check the table for the feature you need, such as streaming, and then test.

When you evaluate vendors, ask four specific questions:

1. Which Arabic locales or dialect modes are supported for streaming (real-time) recognition, not only batch transcription?
2. Is there a mode for mixed Arabic and English speech, or must you choose one language per call?
3. Can you supply custom vocabulary, such as doctor names, department names and drug brand names?
4. Does the text-to-speech offer a Gulf-leaning voice, and does it handle diacritics-free text?

Treat the answers as claims to be tested, not facts.

Code-switching: the normal case, not the edge case

A Gulf caller may say a sentence in Arabic with English nouns inside it: the department, the doctor's title, the word "appointment" or "MRI". Staff and callers who work in English-medium clinical settings do this constantly. The recogniser must cope with it, because the common failure is predictable: the system is set to Arabic, hears an English word, and outputs the nearest Arabic-sounding word. The agent then misreads the intent.

Mitigations, in order of effectiveness:

• Use a recogniser mode that supports multiple languages in one utterance, if your vendor has one. Test it on real mixed sentences.
• Add vocabulary hints for English medical terms and the names of your own doctors and departments, in both Arabic and Latin script.
• Normalise at the output layer. Maintain a mapping from common mis-transcriptions to the intended terms (for example, how a name written in Arabic script maps to the doctor record). Review it regularly.
• Confirm names by readback before booking, and keep the original audio for audit.
• Let the caller switch. If the first reply is in English, the agent should follow into English without friction, and the language field should be stored for the handoff to a human.

Do not force a language menu at the start of the call if you can avoid it. "For Arabic, press 1" loses callers and does not fit how mixed speakers talk. Detect the language from the first utterance and offer a change.

Numbers, dates and times: where bookings go wrong

A booking system is only as good as the digits it captures. Arabic raises specific issues.

Digit forms. Arabic can be written with Western digits (0 to 9) or Arabic-Indic digits, and the two appear in different contexts. Make sure your pipeline normalises both to a single internal format before storing a phone number or date.

Spoken numbers. In Arabic, number words change form with gender and grammatical case, and compound numbers are built differently from English. A recogniser may output a number as words or as digits depending on the mode. Choose a mode that gives digits for phone numbers and test it.

Phone numbers. Callers read them in groups, or in pairs, or digit by digit, and the pattern differs between speakers. The reliable approach is to have the agent read back the number in digit groups slowly, then ask for a yes or no. For the UAE, remember country code variations; accept a number with or without the leading zero or +971, and normalise to one format.

Dates. The Gregorian calendar is used for appointments, but callers may refer to Hijri dates or to days relative to a festival. The agent should use weekdays plus a Gregorian date in readback, for example "Tuesday, the 14th", and avoid relative phrases such as "next week" in a confirmation.

Times. Twelve-hour speech with morning and evening markers is common, but so is a reference to prayer times or to shift hours. Confirm the time explicitly, and avoid assuming 24-hour phrasing.

Emirates ID and other identifiers. Do not read these aloud on a recorded call unless the process truly needs them. Apply the principle of collecting the minimum, and follow the data rules that apply to your emirate and facility.

Readback script pattern

A pattern that works in any language:

1. The agent states what it understood: department, doctor if any, date, time.
2. It reads the phone number in digit groups.
3. It asks one yes or no confirmation.
4. If the caller says no, it asks which item is wrong rather than restarting.

Write this as a template per language, and have native speakers check that the digit grouping sounds natural. Arabic speakers often prefer groups of two or three, but verify with your own callers.

Voice and register

Test several TTS voices with staff and a few real patients. Common complaints about synthetic Arabic are flat intonation, mispronounced names and unnatural stress on English loanwords. Pronunciation dictionaries, where your vendor offers them, are worth building for doctor names, hospital names and common medical terms. Keep replies short; long Arabic sentences are harder to follow on a phone line.

Think about formality too. A reception call in the Gulf has conventions of greeting and politeness that a translated English script will miss. A native-speaking copywriter, not a translation tool, should write the scripts. The same principle applies to written content, and our article on Arabic SEO for clinics explains why Gulf Arabic is researched, not translated.

Test-call protocol: dialect test calls

Because we publish no benchmark, here is the method for producing your own.

1. Recruit speakers. Aim for at least five native speakers per dialect you expect: Gulf (split across several countries if you serve more than one), Levantine and Egyptian, plus bilingual speakers who mix English.
2. Write scenarios. Ten scripts covering booking, rescheduling, doctor enquiry, fee enquiry and directions, each in plain Arabic and in a mixed Arabic-English version.
3. Include digit-heavy tasks: reading a phone number, a date of birth and an appointment time.
4. Record results per call: correct intent, correct slots, number accuracy, readback success, whether a human had to step in.
5. Score by dialect and by mixed versus pure. Differences between groups are your findings.
6. Have native listeners rate the voice for naturalness and politeness on a simple three-level scale.

Make sure each volunteer consents to being recorded, and do not use real patient data in the scripts. Repeat the protocol after every vendor model update.

Regulation and privacy

Voice calls in the Gulf involve personal and health data, and each country and emirate has its own rules on health information, data location and advertising. We have not set out those rules here. Before you deploy, have local counsel confirm recording, consent, storage location and any health-data requirements, and check the claims your agent makes about services against the local advertising rules. Our pages on healthcare marketing in the UAE and the UAE locations hub cover the marketing context.

Where to start

• Decide the output register with native speakers before choosing a voice.
• Ask each vendor the four recogniser questions above, in writing.
• Build custom vocabulary for your doctors, departments and common English terms.
• Normalise digits and phone formats in one place.
• Run the dialect test-call protocol and keep the scoring sheet.
• Make English a smooth fallback, not a separate menu.

For a platform that routes enquiries in the caller's preferred language, see AiTelz, and for the wider service, our AI voice assistant page.