South Asia · Pakistan
Urdu conversational speech dataset
1,000 hours of unscripted Urdu conversation between pairs of native speakers, 100 vetted speakers in all, recorded in Pakistan. Every recording is dual-channel, one speaker per channel, and comes with a time-aligned, human-verified transcript, speaker metadata (age, gender, region and recording device) and signed consent records, under a commercial license for ASR, TTS and voice AI training.
The video couldn’t start. Try playing it again or open the video directly.
Dataset specifications
| Language | Urdu |
|---|---|
| Recorded in | Pakistan |
| Speech type | Spontaneous two-person conversations |
| Volume | 1,000 hours |
| Speakers | 100 vetted native speakers |
| Audio | Dual-channel (one channel per speaker) |
| Transcripts | Time-aligned, human-verified |
| Metadata | Age, gender, region, recording device |
| License | Commercial (training & evaluation) |
| Price | $65 per hour of audio |
Need more hours, another dialect or your own recording setup? We collect speech data to spec.
Built for
For teams training or testing speech models that have to cope with real Urdu speech:
- Automatic speech recognition (ASR / STT)
- Text-to-speech and voice synthesis prosody
- Voice agents and IVR for Pakistan
- Speaker diarization and turn-taking models
- Speech LLM and audio-understanding fine-tuning
Why conversational data?
Models trained on read speech lose accuracy on spontaneous audio, because hesitations, overlapping turns and natural speaking rates never made it into their training data. This dataset fills that gap for Urdu. Our guide to conversational speech data explains why it matters.
Free samples
Audit the Urdu data yourself
Leave your details and the Spirelight team will email representative audio, matching transcripts and the metadata schema within 48 hours. Check recording quality, transcript accuracy and the speaker mix before you commit to anything.
FAQ
Frequently asked questions
What is in the Urdu speech dataset?
1,000 hours of unscripted conversation between pairs of native speakers, 100 vetted speakers in all, recorded in Pakistan. Each conversation is delivered as dual-channel audio with a time-aligned transcript, speaker metadata (age, gender, region and recording device) and the consent records behind it.
What makes Urdu hard for speech recognition?
Urdu uses the Perso-Arabic script, normally without short vowels, so a written word can often be read more than one way. Its grammar is largely shared with Hindi while its vocabulary leans on Persian and Arabic, and Pakistani speakers mix in a good deal of English and Punjabi.
Is the Urdu dataset licensed for commercial use?
Yes. The license covers commercial training and evaluation of ASR, TTS and other voice AI models. Every speaker signed explicit consent to commercial use, and Spirelight shares the full consent records for your legal review.
What does the Urdu speech dataset cost?
$65 per hour of audio. Buying the full dataset, or several languages together, qualifies for a volume discount: ask for a quote.
Can I check samples before I buy?
Yes. Request samples on this page and the Spirelight team emails representative audio, matching transcripts and the metadata schema within 48 hours, so your team can audit the data before committing.