Urdu · national language of Pakistan · Nastaliq script · ~230 million speakers worldwide
Urdu AI training data: speech, text, OCR and RLHF
Urdu has well over 200 million speakers, is mutually intelligible with spoken Hindi, and is still badly served by open corpora: read speech, news text, no Roman Urdu, no telephony, no code-switching. We collect the Urdu production systems actually hear — nationwide, to quota, consented per record and licensed for training.
At a glance
Urdu training data: the facts a buyer needs first
What decides whether an Urdu programme succeeds, before anyone writes a statement of work.
| Speakers | ~80 million L1+L2 in Pakistan (national language and lingua franca); ~70 million in India; large diaspora in the UK, US, Canada, Gulf and Australia. Spoken Urdu and Hindi are largely mutually intelligible |
|---|---|
| Script | Perso-Arabic in the Nastaliq style — slanted, stacked ligatures that break OCR trained on Naskh Arabic. Also written informally as Roman Urdu in Latin script across chat, search and social media |
| Varieties we field | Standard (Karachi, Islamabad), Lahori, Dakhini-influenced Karachi Urdu, and Urdu–English code-switched urban speech; regional-accented Urdu from L2 speakers (Punjabi, Pashto, Sindhi L1) to quota |
| Where we collect | Nationwide — deepest pools in Karachi, Lahore, Islamabad–Rawalpindi, Faisalabad, Multan, Peshawar and Hyderabad, plus rural districts in every province |
| Capability tier | Deep speech · Deep text · Deep image/OCR |
| Typical pilot | 50–200 hours of speech, or 25k–100k text items, 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
- Conversational Urdu Speech — spontaneous, diarized, 200–500 hrs first delivery
- Pakistani Telephony Corpus — 8 kHz call-centre Urdu
- Nastaliq & Handwriting OCR — forms, receipts, prescriptions
- Native Urdu Instruction & SFT Corpus
- Urdu–English Parallel Corpus — six domains
- RLHF Preference Pairs and Toxicity Lexicon
Built for
- Fine-tuning Whisper, MMS, wav2vec 2.0 or in-house ASR for conversational and telephony Urdu
- Urdu voice agents, IVR and call analytics for banks, telcos and government
- Urdu LLM fine-tuning, evaluation and safety alignment
- Nastaliq OCR for KYC documents, prescriptions, ledgers and archives
- Urdu↔English and Urdu↔Hindi speech translation
Why models fail
The four things that break Urdu models
Open Urdu corpora are read speech and news text. Production Urdu is none of those things.
Nastaliq
Urdu is printed in Nastaliq, where letters stack diagonally and word shapes change with context. OCR trained on Naskh Arabic reads it as noise. Ground truth has to be Urdu-native, with reading order and ligature-aware segmentation. See why Urdu OCR is harder than Arabic OCR.
Code-switching
Urban Urdu switches into English every few words — nouns, verbs, whole clauses. ASR trained on monolingual Urdu misses half the content. Transcription conventions must state how English spans, hybrid verbs and Roman-script tokens are written. See how to fix code-switching in ASR.
Register and Roman Urdu
Formal Urdu is Persianised; conversational Urdu is not. Most typed Urdu online is Roman Urdu in Latin script with no standard spelling. Text models trained on news Urdu fail on how people actually write, and search, chat and moderation systems need Roman Urdu explicitly.
Hindi confusion
Spoken Urdu and Hindi overlap heavily, so multilingual models mislabel Urdu audio as Hindi and produce Devanagari output. Urdu data transfers measurably into Hindi ASR — and vice versa — but only if language ID and script policy are handled on purpose. See what transfers and what does not.
What we collect
Urdu speech, text, image and annotation data
Every stream is quota-controlled: demographic and acoustic spec approved by you, recruitment in the districts where speakers live, out-of-quota submissions rejected even when technically clean.
Speech for ASR
Spontaneous two-party conversation, 8 kHz telephony, read speech, wake words and commands, far-field and in-vehicle audio. Quotas across age, gender, urban/rural, region, L1 background and device. Verbatim transcripts with a written code-switching convention, diarization and noise tags.
Speech for TTS
Studio single- and multi-speaker Urdu voices, phonetically balanced scripts, 44.1/48 kHz, pronunciation lexicon with Perso-Arabic loanword handling. 10–40 hours per voice; see how much a TTS voice needs.
Text, SFT and RLHF
Native-authored instruction data and multi-turn dialogues in Urdu script and Roman Urdu, preference rankings by calibrated native raters, severity-graded toxicity data, and Urdu–English parallel text across law, medicine, finance, news, government and colloquial registers.
OCR, documents and vision
Printed Nastaliq and handwritten forms, receipts, prescriptions and ledgers with field-level ground truth; Urdu-labelled street, retail and document imagery; consented face and liveness capture. See image & video data.
Ready-made
Catalog datasets that include Urdu
Each ships with a datasheet, per-record consent artefacts and a free sample pack. Subsets by language are available on request.
Conversational Urdu Speech
Spontaneous, diarized, 4 age bands, urban/rural
SpeechPakistani Telephony Corpus
8 kHz call-center audio in Urdu, Punjabi, Sindhi, Pashto
ImageNastaliq & Handwriting OCR
Printed and handwritten forms, receipts, prescriptions
TextNative Urdu Instruction & SFT Corpus
Human-authored, multi-turn, culturally grounded
TextUrdu–English Parallel Corpus
Human-translated, six domains
TextRLHF Preference Pairs
Native rater rankings with rationales
Specification
Writing a Urdu spec that survives QA
Most Urdu projects that disappoint were under-specified. These decisions belong in writing before recording starts.
Fix the script policy
Urdu script, Roman Urdu, or both; how English tokens are written (source script vs transliteration); whether diacritics (aerab) are transcribed.
Write the code-switching convention
Language-span tags or not; treatment of English loanwords that are now Urdu; hybrid verbs like “download karna”. Two pages, agreed before QA.
State the register target
Conversational, formal, or a stated mix; whether Persianised vocabulary is in or out of scope for TTS and SFT.
Set quota axes
Age bands, gender, urban/rural, province, L1 background (native Urdu vs Punjabi/Pashto/Sindhi L1), device class — and how strict.
Define acoustic conditions
Telephony 8 kHz vs wideband; far-field and in-vehicle share; market and traffic noise; handset diversity.
Set the acceptance test
Held-out set from the same quota, WER target against a reference transcript, inter-annotator agreement floor on a 5% double-transcribed sample.
Who buys Urdu data
Urdu is bought in Karachi, Dubai, London, Toronto and San Francisco
Domestic demand from Pakistan's banks, telcos and government; diaspora demand from the UK, US, Canada, Gulf and Australia; and frontier-lab demand for South Asian coverage and evaluation.
- Voice agents and IVR — banks, telcos, utilities and government helplines in Pakistan; NHS, council and utility services in the UK where Urdu is a top non-English language
- LLM localisation and safety — Urdu SFT, RLHF and red-team data for assistants that must handle Urdu script and Roman Urdu
- KYC and document AI — Nastaliq OCR for CNIC, utility bills, prescriptions and ledgers; consented face and liveness data
- Speech translation and interpreting — healthcare, legal and immigration settings in Canada, the US, the UK and Australia
- Hindi systems — Urdu speech as an acoustic supplement for Hindi ASR, with measured transfer
Answers
Questions buyers ask about Urdu data
Ready-made: Conversational Urdu Speech (200–500 hours first delivery) and the Urdu portion of the Pakistani Telephony Corpus are licensable today with a free sample in one business day. Custom collection scales to several thousand hours over a programme from standing panels in every major city.
Yes. Roman Urdu text (chat, search, social, reviews) and Roman Urdu transcripts of speech are available, with a stated normalisation convention because Roman Urdu has no standard spelling. Most production text systems for Pakistan need Roman Urdu more than Urdu script.
Partly. Spoken Urdu and Hindi overlap heavily, so Urdu speech measurably improves Hindi ASR and vice versa. Text does not transfer because the scripts differ, and vocabulary diverges in formal registers. Our transfer study documents the WER deltas; ask for it.
Urdu is printed in Nastaliq, not Naskh. Letters stack diagonally, baselines slope and ligature shapes vary with context, so segmentation trained on Arabic fails. Our OCR sets carry Urdu-native ground truth with reading order and field labels.
Yes. Most of our licensing goes to teams outside Pakistan. English-language contracts under an agreed governing law, USD invoicing by default, W-8BEN-E for US payers, and delivery to your cloud region with a DPA. See Where we deliver.
Every contributor signs or records a plain-language Urdu consent granting perpetual, worldwide, sublicensable training rights including derivative models. The consent version and contributor pseudonym travel with each record, so counsel can trace any row to its signature.
Tell us what your Urdu model is getting wrong
Two lines — the task, the register, the acoustic conditions. You get a scoped collection design, a free sample and an honest timeline back, not a discovery-call funnel.
Average first response: under 6 business hours. NDAs signed same day.