One Caller Speaks Swahili, the Next Malayalam. Then What?

A caller rings your front desk and opens in Swahili. The next one wants an appointment in Malayalam, and the one after that switches from Omani Arabic to English halfway through a sentence. How your line copes with that decides who actually gets served and how much it quietly costs you, whether it relies on language detection, on a colleague down the corridor, or on callers who simply give up.

The languages already on your line

More than four in ten people living in Oman are expatriates, according to the population clock of the National Centre for Statistics and Information. NCSI figures reported for June 2025 counted about 621,000 Bangladeshi, 508,000 Indian and 312,000 Pakistani expatriate workers, plus around 26,000 from Tanzania. Add families here who grew up speaking Swahili at home, plus tourists in high season, and one phone line hears a lot of languages.

• Arabic: mostly Omani Arabic, plus Egyptian, Sudanese and other dialects.
• English: the shared second language of most desks and most tourists.
• Hindi and Urdu: close enough in everyday speech that many callers move between them.
• Bengali: spoken by one of the largest communities in the country.
• Malayalam and Tamil: from South India and Sri Lanka, common in healthcare, retail and services.
• Swahili: from East African families and residents.

Most of these callers can manage some English or Arabic. But managing is not the same as explaining a billing dispute, or describing chest pain to a nurse.

What "press 2 for English" does to everyone else

A typical IVR (interactive voice response: the recorded menu that asks callers to press numbers) offers two languages. A Bengali-speaking customer chasing a delivery presses 2, hears a fast English menu, guesses, and lands in the wrong queue. A mother who is more comfortable in Malayalam gets through to a polyclinic, then spends four minutes spelling her son's name while the operator guesses.

On the operator floor, the usual fix is human: "Who speaks Tamil? Is Ravi on shift?" The caller waits on hold, the colleague leaves their own queue, and often the call ends before anyone is found.

Callers who struggle with the language of your phone line rarely complain: they quietly stop calling, and your reports never show it.

What the language gap costs a busy desk

Take a composite example. Badriya coordinates the call desk for a polyclinic group in Muscat with three branches and one main number. Here is her month, worked out so you can swap in your own figures.

Hiring someone who speaks every language is not realistic. What closes most of that gap is a different way of answering the phone, and it helps to understand how it works before anyone sells it to you.

How language detection works, in plain words

Language detection means the software works out which language a caller is speaking from the sound of their voice. It is the same trick you use on a bus when you can tell Malayalam from Arabic without understanding a word. Modern AI voice agents do it in five steps.

1. Listen first. The agent opens with a short greeting and lets the caller talk. A second or two of speech is usually enough for a first guess.
2. Check its confidence. If the guess is strong, it carries on. If not, it asks what a good operator would: "Arabic, English, or another language?"
3. Switch completely. A caller who says "Habari, nataka kujua kama daktari yupo kesho" (Hello, I want to know if the doctor is in tomorrow) hears Swahili back.
4. Keep listening. A caller may start with "Doctor-inte appointment venam" in Malayalam, then read out a number in English. A good agent follows each switch instead of sticking with the first language it heard.
5. Write it down for your team. Whatever language the call was in, the summary, booking and ticket arrive in Arabic or English, so the person following up can read them.

Operations leads tend to overlook step five. Your caller is served in Bengali, and your supervisor still reads every call. We followed one such call from start to finish in The Caller Who Switched Languages Three Times.

Where it still stumbles

Arabic is the first test, not the easy one. If a system hears "abgha maw'id baacher" (I want an appointment tomorrow) and answers in stiff textbook Arabic, or mishears the dialect altogether, it has detected the language and still missed the caller. Other weak spots are worth knowing too:

• Not every language is equally strong. AI developers have put most of their effort into Arabic and English. Quality in Swahili, Bengali, Malayalam and Tamil varies a lot from one system to the next.
• Names, numbers and addresses are where errors hide. A misheard digit in a civil ID or plate number does more harm than a clumsy sentence.
• Mixed speech, such as Malayalam threaded with English or Arabic mixed with English, trips up weaker systems.
• Sensitive moments, such as a complaint or a medical symptom, should go to a person. A good set-up hands over the call with notes attached, using escalation rules you decide in advance.

What this means for you

You do not need an operator for every language. You need three things: a first answer in the caller's own language, names and numbers captured exactly, and a clean handover to a human with notes your team can read.

Start by replacing "press 1, press 2" with a greeting that lets callers simply talk. Then decide which languages the agent handles end to end, and which ones it only greets, takes details in and passes to a person. Keep your multilingual staff for calls that need judgement, not for spelling names back.

That is the plan Badriya is taking to her managers: the agent takes the first minute of every call in the caller's language, and her Tamil- and Bengali-speaking colleagues take the calls where a human voice matters. Agentic Customer Care by CustomerCare.OM works this way: it answers in the caller's language, books the slot or raises the ticket, and leaves the team a summary in Arabic or English.

The bottom line

Your phone line is already multilingual, whether or not your menu admits it. Language detection lets every caller start in the language they think in, while your team keeps reading every call in its own. Run the ten-call test first, trust only the languages that pass, and send the rest to the people on your floor who speak them.