Urdu Speech Evaluation Set
A balanced 9.7-hour test set for Urdu speech recognition and synthesis: 10 speakers (5 female, 5 male) from all four provinces of Pakistan, at most one hour each, stratified across 25 domains. 3,178 clips, each with its sentence in Urdu script, and a fixed speaker-disjoint dev/test split (splits.tsv).
These recordings are a subset of the Urdu Read-Speech TTS Corpus (~502 h, 21 speakers).
Language and script
Urdu (ISO 639-3: urd), the national language of Pakistan. Text is in the Perso-Arabic Urdu script (Unicode Arabic block).
How it was collected
- Read speech recorded in Pakistan between July and September 2026. Each speaker read sentences shown in the SABRe recording app, grouped by domain.
- Collected for Mozilla Data Collective by MirasAI (mirasai.net).
- Two recording setups: most speakers used a Boya K3 desktop microphone in a local studio setup; Speaker-15, Speaker-17 were recorded in a professional studio (Audio-Technica AT2020) at a lower level (about 13 dB quieter), left as recorded.
- All speakers agreed to take part and were paid. Consent forms are available on request.
- Sentence sources: over 90% from Kaleem Art Press corpora (Kaleem Art Press is MirasAI's on-ground team), under 10% generated with Google AI Studio.
- Speakers are identified by ID only; no names are published.
Contents
| |
|---|
| Hours | 9.68 |
| Clips | 3,178 |
| Speakers | 10 |
| Domains | 25 |
| Unique sentences | 801 |
Splits (speaker-disjoint)
| Split | Speakers | Hours | Clips |
|---|
| dev | Speaker-20, Speaker-06 | 1.94 | 564 |
| test | Speaker-03, Speaker-05, Speaker-15, Speaker-17, Speaker-01, Speaker-08, Speaker-09, Speaker-21 | 7.74 | 2,614 |
Domains
| Domain | Hours | Clips | Speakers |
|---|
| Literature | 1.1 | 375 | 7 |
| Science and Technology | 1.0 | 354 | 9 |
| Religion | 0.9 | 338 | 8 |
| Health | 0.9 | 314 | 7 |
| Agriculture | 0.6 | 140 | 5 |
| Instructional | 0.5 | 113 | 4 |
| Social Media | 0.4 | 150 | 7 |
| Culture |
Speakers
| Speaker | Gender | Age | Region | Hours | Clips | Domains |
|---|
| Speaker-03 | Female | 30-39 | Multan, Punjab | 1.0 | 441 | 10 |
| Speaker-05 | Female | 50-59 | Khanewal, Punjab | 1.0 | 327 | 16 |
| Speaker-15 | Female | 18-19 | Faisalabad, Punjab | 1.0 | 244 | 7 |
| Speaker-17 | Female | 30-39 | Quetta, Balochistan | 1.0 |
Audio
All clips are 44.1 kHz mono 16-bit PCM WAV. Clip length: 2.4 s to 29.6 s, median 9.6 s.
Files
Speaker-NN/
mapping.tsv audio_filename, key, domain, sentence (tab-separated, UTF-8)
clips/ one WAV per row, named .wav
speaker_x_domain.csv hours per speaker and domain
splits.tsv speaker, split (dev/test), gender, age, region, province, clips, hours
README.md
LICENCE.md
key is the SHA-256 hash of the sentence text, so the same sentence has the same key for every speaker.
Processing
- Same processing as the full corpus: domain labels merged into 25 domains, one recording per sentence per speaker, every row has its audio, stereo and 48 kHz recordings converted to 44.1 kHz mono.
- Clips of 1-30 s only. The sentences in this set are withheld from the other Urdu subsets we sell, so models trained on those can be tested on this set without sentence overlap.
License and access
MDC Data Licence Agreement 1.0 (commercial; see LICENCE.md). Access is restricted.
Contact
Mozilla Data Collective, via this dataset's page on mozilladatacollective.com
