Serving Customers in Their Own Language, Properly

The claim and the reality

Almost every voice platform advertises support for a long list of languages. The list is usually accurate in the narrow sense that the system will produce output in those languages. Whether it does so well enough to put in front of a paying customer is a different question.

Everything below came from listening to real production calls, not from testing. None of it appeared in advance.

People mix languages in one sentence

A caller says an Urdu sentence with English technical words in the middle of it. Or Arabic with English product names. This is completely normal speech across South Asia, the Gulf and diaspora communities — and it defeats systems that pick one language at the start of a call and commit to it.

If your customers do this, test for it explicitly. Ask a vendor to demonstrate a mixed-language call rather than a clean one.

Right-to-left scripts break in places you would not check

Arabic and Urdu need correct handling at every stage: recognition, the model, and the text that finally lands in your helpdesk. We have seen transcripts arrive in the wrong script entirely on short replies — the meaning survived, but the stored record was wrong, which matters once a human reads it later.

Replies get cut off mid-word

A response limit that is generous for English can be too small for Urdu or Arabic, because those scripts cost several times more tokens per word. The symptom is an agent that stops mid-sentence, which reads as a fault rather than a limit.

Local names are mis-heard

Place names, street names and product names specific to your market are exactly what a general-purpose recogniser has not been trained on. They are also what customers say most often. This is usually fixable by biasing the system toward a word list, but only if someone notices it is happening.

The voice can be fluent and still sound wrong

A voice can be grammatically perfect and unmistakably foreign to a local ear. A British-accented voice reading Urdu is technically correct and lands badly with the customer it is meant to serve. Always listen to a candidate voice with someone who speaks the language natively, before launch.

What to ask a vendor

• Can you demonstrate a call where the customer switches language mid-sentence?
• What does the stored transcript look like for a right-to-left language?
• Which accent does the voice have, and can we hear it before we commit?
• How do you handle names specific to our market?

These questions are easy to answer if the vendor has genuinely done the work, and awkward if the language support is a checkbox.

If you need to serve customers in a language your current platform handles badly, tell us which languages your customers actually call in.