Saraiki is spoken by tens of millions of people in southern Punjab, yet almost no public speech data exists for it. We test your ASR, voice agent or TTS on consented Saraiki speech recorded where the language is actually spoken.
Models trained on Urdu or Hindi often replace Saraiki words with the nearest Urdu word, and the meaning changes.
Everyday Saraiki verbs and nouns differ from Urdu, so Urdu-trained models substitute them.
Saraiki has implosive consonants that Urdu does not, which ASR often mishears.
Very little transcribed Saraiki speech exists publicly, so most models have never heard it.
We split results by variety, recording condition and speaker group so you see exactly where the model breaks.
| Area | What we measure | Example (illustrative) |
|---|---|---|
| Dialect substitution | Saraiki word replaced by Urdu or Hindi word | wanjna → jana |
| Implosive sounds | Errors on sounds Urdu lacks | ɓ, ɗ, ʄ, ɠ heard as plain consonants |
| Variety gap | Accuracy by Multani, Derawi, Riasti | per-variety WER and CER |
| Rural audio | Speech in homes, fields and markets | noise type logged per clip |
| Meaning-critical | Transcript close, meaning wrong | negation or time word dropped |
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 Saraiki sample pack is being recorded now. Request a pilot and we send the specimen report and file format.
Illustrative example of our format. Live consented samples publish at launch.
A fixed-scope Saraiki pilot with a full error report. You keep the findings.