Clean studio Urdu is easy to find. The Urdu your users actually speak is not: regional accents from Saraiki and Punjabi speakers, and English words inside almost every sentence. We record it with consent and test your model on it.
A model can score well on clean read Urdu and still fail on a caller from Multan who mixes in English.
Most Pakistanis speak Urdu as a second language, with Saraiki, Punjabi or Pashto accents.
Words like account, reschedule and balance appear inside Urdu sentences.
Phone lines, markets and homes, not quiet studios.
Results are split by accent, code-switch density and recording condition.
| Area | What we measure | Example (illustrative) |
|---|---|---|
| Code-switch failure | English words misspelt or dropped | reschedule → re shedule |
| Accent gap | Accuracy by speaker's first language | Saraiki-accented vs Punjabi-accented WER |
| Roman Urdu output | Spelling consistency and normalisation | kya / kia / kiya |
| Numbers and dates | Amounts, dates and times spoken in Urdu or English | dhai hazar → 2,500 |
| Meaning-critical | Transcript close, meaning wrong | kal (tomorrow) ↔ aj (today) |
No long contract. Test one model on real regional speech, then decide.
| Pilot | Scope | Price |
|---|---|---|
| Starter test | 100 utterances · 1 language · WER/CER + error classes | from $499 |
| Regional pilot | 300 utterances · up to 3 languages · full written report | from $999 |
| Custom | Your domain, hours, conditions and schedule | Quoted after scoping |
Fixed-scope pilots. You keep the report and the per-utterance results. Your test set stays private and is never resold.
Our first consented Pakistani Urdu sample pack is being recorded now. Custom Urdu datasets are also available. Request a pilot and we send the specimen report and file format.
Illustrative example of our format. Live consented samples publish at launch.
Evaluation pilot or custom dataset, scoped to your product.