Urdu and Roman Urdu in software: what works today
Customers write "order kab ayega?" as often as "when will my order arrive?". Language models now handle Roman Urdu and Urdu far better than they did, with a few gaps worth planning around.
Messages to Pakistani businesses arrive in a mix of languages. English, Urdu script and Roman Urdu, often in the same sentence: "Mera order abhi tak deliver nahi hua, tracking number send karein". Software that only understands English misses a large share of what customers say.
What works well now
- Understanding intent: modern language models recognise what a Roman Urdu or Urdu message is asking for, even with mixed English words.
- Replying in the customer's language: the same assistant can answer in English, Urdu or Roman Urdu depending on how the customer wrote.
- Sorting messages: complaints, order questions, price questions and sales leads can be separated reliably.
- Summaries: a long Roman Urdu voice-note transcript or chat can be summarised in English for the team.
Where the gaps are
- Spelling: Roman Urdu has no standard spelling. "Kab", "kb" and "kub" all mean the same thing, which hurts simple keyword search.
- Urdu text in scanned documents: reading printed Urdu from images is still much weaker than reading English.
- Speech: transcription of Urdu voice notes is usable but makes more mistakes with names, places and fast speech.
- Smaller models: compact models that run on a modest server handle Roman Urdu less well than the large hosted ones.
Many spellings, one meaning. Match on meaning, not on exact words.Practical choices
Common Urdu and Roman Urdu tasks and the approach that tends to work.| Task | Approach | Note |
|---|
| Search a help centre | Meaning-based search, not keywords | Handles spelling variants |
| Route WhatsApp messages | Language model classification | Test on real messages first |
| Reply to customers | Model replies in the customer's language | Keep numbers from lookups |
| Transcribe voice notes | Speech model, then a summary | Check names and amounts |
| Read Urdu scanned forms | Specialised OCR plus review | Expect manual checks |
Test with your own messages
Building the test set
The best test set is a few hundred real, anonymised messages from your own customers. They carry the spellings, slang and product names that general benchmarks never include. We measure on those before promising anything.
Customers will not switch to English for your software. The software has to meet them in the language they type.
Frequently asked questions
Can the bot reply in Urdu script?
Yes. It can reply in Urdu script, Roman Urdu or English, usually matching the way the customer wrote. Some businesses prefer always replying in one language, which is a simple setting.
Do we need to label thousands of messages?
Usually not anymore. Large language models understand intent with a short description and a few examples. Labelled messages are mainly needed to test accuracy.
What about Punjabi, Sindhi or Pashto?
Support is weaker and uneven. We test on real samples before relying on it, and keep a person in the loop for those conversations.
Can it understand voice notes in Urdu?
Yes, with a speech model that transcribes the voice note first. Accuracy is good for clear speech and weaker with noise, names and fast talking, so important details are checked.