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.
Why multilingual voice agents fail at the switch
In Pakistan, a customer might say "Mera balance check kar dein, aur last transaction bhi bata dein." Half of it is English. Speech models that expect one language per utterance often misspell the English words, split them, or drop them entirely. The agent then misses exactly the words that carry the intent: balance, transaction.
Designing realistic conversational recordings
- Use real scenarios: banking, telecom, delivery, appointments.
- Let speakers phrase requests naturally instead of reading a script.
- Include different switching patterns: single English words, full English phrases, and English numbers inside Urdu.
- Record on phones as well as microphones, with real background noise.
Language-switch labels and reference transcripts
Each English span should be written in English script and tagged with its language. That lets you measure errors separately for the Urdu parts and the English parts, and see exactly where your model loses accuracy.
Stratified evaluation, WER and meaning-critical failures
Report WER separately for utterances with and without code-switching, and for Urdu versus English spans. Then add a meaning-critical check: did the agent understand the request? An utterance with one small spelling error can still change the intent, for example when last is misheard and the agent returns the wrong transaction.
Annotation schema
A simple, portable format works best. We deliver JSON Lines with one record per utterance:
{"id":"URD-0311","text":"Mera balance check kar dein",
"spans":[{"text":"balance","lang":"en"},{"text":"check","lang":"en"}],
"intent":"balance_inquiry","condition":"phone_call","speaker_l1":"Punjabi"}
Want this applied to your own audio? Send 10 minutes and your guidelines, and we will return a labelled sample.