Urdu · Datasets · 9 min read
Urdu Speech Datasets for ASR: What AI Teams Should Evaluate
Urdu is spoken by well over 200 million people, yet it remains thin in web data. Here is what to check before you train or test an Urdu speech model.
Who speaks Saraiki
Saraiki is spoken across southern Punjab and nearby regions, with major centres in Multan, Bahawalpur, Dera Ghazi Khan and Rahim Yar Khan. It has several recognised varieties, including Multani, Riasti (around Bahawalpur), Derawi (around Dera Ghazi Khan) and Thali (around Layyah). Tens of millions of people use it at home, in markets and on the phone every day.
Why models mishear it
Most speech models learn from audio and text dominated by Urdu and Hindi. Saraiki shares many words with them, but its distinctive pronouns, verbs and endings are rare in training data. The model "corrects" what it hears toward the closest Urdu form.
Because written Saraiki is so scarce online, text-based language models cannot fix these errors on their own. The only reliable solution is real, verified Saraiki speech with native transcripts.
Orthography decisions
Saraiki can be written in Shahmukhi (Perso-Arabic script, with extra letters for implosive sounds) or in Latin script, and everyday spelling is inconsistent. Before transcription starts, decide which script and which spelling rules to use, then apply them everywhere. Otherwise the same word appears in three spellings and your error rates become meaningless.
Ethical native-speaker recruitment
- Recruit through trusted local coordinators, not anonymous online panels.
- Explain the purpose in Saraiki: the recordings may train or test AI systems.
- Pay speakers fairly for their time, on the day.
- Record adults only, and never collect ID numbers or personal details in the audio.
- Give speakers a way to ask for future use of their recordings to stop.
Recording and review protocol
Record in the places people really talk, such as courtyards, markets and phone calls, and log the condition for every clip. Each recording is transcribed by a native Saraiki speaker and checked by a second reviewer. Disagreements go to a senior linguist, and the decision is recorded.
Commissioning a Saraiki dataset
Start with a scoped pilot: about 100 utterances across two or three varieties and conditions, with a full error report on your model. That tells you quickly where your model breaks and how much data you actually need.