Roman Urdu · Urdu written in Latin script · how Pakistan actually types
Roman Urdu datasets: the Urdu your text models never see
Most Urdu typed online is not in Urdu script. It is Roman Urdu — Latin letters, no standard spelling, mixed with English — in chat, search, reviews, support tickets and social media. Open corpora are almost entirely Nastaliq news text, so models fail on what users actually write. We produce Roman Urdu text, labels and speech transcripts to a stated normalisation convention.
At a glance
Roman Urdu training data: the facts a buyer needs first
What a buyer needs to know before commissioning Roman Urdu data.
| What it is | Urdu (and often Punjabi) written in the Latin alphabet, as on phone keyboards and in messaging apps. Not a transliteration standard — a set of habits, with the same word spelled several ways (mein / main / mai / mn) |
|---|---|
| Who writes it | The large majority of Pakistan's 110M+ social-media users, plus the diaspora in the UK, US, Canada, Gulf and Australia messaging family and businesses |
| Where it matters | Search queries, chat and voice-assistant text, customer-support tickets, product reviews, comments and moderation, SMS and WhatsApp business messaging, ad copy |
| The core problem | Spelling variation, heavy English mixing and no public corpus of everyday register. Tokenizers and classifiers trained on English or Nastaliq Urdu treat it as noise |
| Capability tier | Deep text · Deep labels · Deep Roman-script speech transcripts |
| Typical pilot | 25k–100k text items or 50–200 hours of speech with Roman transcripts, in 3–5 weeks; creditable against production |
| Licence | Perpetual, worldwide, sublicensable; derivative-model rights stated explicitly. Licensing detail |
| Delivery | Your S3, GCS or Azure bucket with manifests, checksums, QA report and datasheet |
Speaker figures are approximate, drawn from the 2023 census and public sources, and include second-language speakers where noted.
What you can license today
- Roman Urdu variants in the Safety & Toxicity Lexicon
- Roman Urdu prompts and responses in the Native Urdu SFT Corpus
- Roman-script transcripts as an option on every Urdu speech corpus
- Custom: chat and support-ticket style corpora, sentiment and intent labels, search-query sets, Roman↔Nastaliq parallel text
Built for
- Sentiment, intent and toxicity classifiers for Pakistani user text
- Search, autocomplete and spelling normalisation for Roman Urdu queries
- Chat assistants and LLM fine-tuning that must read and write Roman Urdu
- Transliteration models Roman Urdu ↔ Urdu script
- Code-mixed Urdu–English NLP
Why models fail
Why models trained on Urdu script fail on Roman Urdu
Roman Urdu is not Urdu with a different font. It is a different distribution.
No standard spelling
The same word appears in five or more spellings, varying by region, age and keyboard. Vocabulary fragments, frequency counts lie, and out-of-vocabulary rates explode unless a normalisation convention is applied and documented.
English everywhere
Roman Urdu text is code-mixed by default: English nouns, verbs and whole clauses in the same sentence, in the same script. Language identification at the token level is a prerequisite for anything downstream.
Register is informal
Abbreviations, elongation for emphasis (bohooot), emoji, and sentence structure that never appears in news Urdu. Models trained on formal text mis-tokenize and mis-classify it.
Open data is tiny and skewed
Public Roman Urdu sets are small, mostly scraped comments, unlabelled or weakly labelled, and legally unclear. Production systems need consented, labelled, register-balanced corpora.
What we collect
Roman Urdu speech, text, image and annotation data
Native authors and raters producing text in the register you specify, labelled to your taxonomy, with the normalisation convention agreed before QA starts.
Authored and curated text
Chat-style dialogues, support tickets, reviews, comments and search queries authored natively in Roman Urdu across domains you specify, balanced by register and region, with consent on every item.
Labels
Sentiment, intent, topic, toxicity and spam labels by three calibrated native raters with agreement reported; token-level language tags for Urdu / English / named entities.
Roman-script speech transcripts
Any Urdu or Punjabi speech corpus delivered with Roman Urdu transcripts (in addition to or instead of Nastaliq) to a stated convention — what ASR for chat-first products needs.
Parallel and normalisation data
Roman Urdu ↔ Urdu-script parallel sentences and spelling-variant clusters (mein / main / mai → canonical) for transliteration and normalisation models.
Ready-made
Catalog datasets with Roman Urdu content
Each ships with a datasheet, per-record consent artefacts and a free sample pack. Subsets by language are available on request.
Safety & Toxicity Lexicon
Roman Urdu spelling variants included
TextNative Urdu Instruction & SFT Corpus
Urdu script and Roman Urdu
TextRLHF Preference Pairs
Native rater rankings
SpeechConversational Urdu Speech
Roman transcripts on request
SpeechPakistani Telephony Corpus
Roman transcripts on request
TextUrdu–English Parallel Corpus
Human-translated, six domains
Specification
Writing a Roman Urdu spec that survives QA
Roman Urdu projects live or die on the normalisation decision. Make it first.
Choose a normalisation policy
Preserve spelling as written (for robustness), normalise to a canonical form (for search), or deliver both with a mapping. Most teams need both.
Fix the English policy
Whether English tokens are tagged, kept as written, or normalised to dictionary spelling; how hybrid verbs (download karna) are tagged.
Define the register and sources
Chat, reviews, tickets, search, social — in stated proportions; authored to spec, never scraped without consent.
Set the label taxonomy
Sentiment scale, intent classes, toxicity categories and severity grades, with definitions and worked examples in Roman Urdu.
Set the acceptance test
Inter-annotator agreement floor per label, out-of-vocabulary rate after normalisation, held-out test set from the same distribution.
Who buys Roman Urdu data
Who buys Roman Urdu data
Anyone whose product receives typed input from Pakistani users — which is most consumer and fintech products in the market, and every platform with a South Asian audience.
- E-commerce and fintech — search, reviews, chat support and fraud text in Roman Urdu
- Social and messaging platforms — moderation, spam and sentiment on Roman Urdu content
- LLM and assistant builders — Roman Urdu understanding and generation, transliteration to script
- Telecom and utilities — SMS and WhatsApp business messaging intent detection
- Research groups — labelled, consented Roman Urdu corpora with documented conventions
Answers
Questions buyers ask about Roman Urdu data
No. Transliteration follows a standard mapping from script to Latin letters. Roman Urdu is how people actually type, with no standard, so the same word has many spellings. We can deliver both, with a mapping between them.
You choose: preserve as written, normalise to a canonical form, or both. The convention is written down, applied by all annotators, and inter-annotator agreement is reported so you can see it holding.
Yes. Any Urdu or Punjabi speech corpus can be transcribed in Roman Urdu, Urdu script, or both, to a stated convention. Chat-first voice products usually want Roman.
No. Text is authored or contributed under consent, or labelled on data you already own and have rights to. Scraped comment corpora carry licensing risk our buyers cannot accept.
Yes. Punjabi is also typed in Latin script in Pakistan and the diaspora; the same authoring, labelling and normalisation process applies.
Yes; most of our licensing goes to teams outside Pakistan. English contracts, USD invoicing, delivery to your cloud region with a DPA. See Where we deliver.
What does your model do with “mjhe ye order cancel krna hai”?
If the answer is “nothing useful”, tell us the task. We come back with a data design, a normalisation proposal and a free labelled sample.
Average first response: under 6 business hours. NDAs signed same day.