Urdu Read-Speech TTS Corpus
About 502 hours of read Urdu speech for text-to-speech: 169,661 clips from 21 native speakers (13 male, 8 female) across 25 domains, each clip paired with its sentence in Urdu script. Recorded in Pakistan, July-September 2026.
Language and script
Urdu (ISO 639-3: urd), an Indo-Aryan language and the national language of Pakistan. Text is in the Perso-Arabic Urdu script.
How it was collected
- Read speech recorded 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; speakers 15-19 were recorded in a professional studio with Logic Pro (Audio-Technica AT2020 cardioid condenser microphone). That batch was recorded at a lower level (about 13 dB quieter) and is left as recorded, so users can normalise it if they need to.
- 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 | 501.83 |
| Clips | 169,661 |
| Speakers | 21 (13 male 301 h, 8 female 201 h) |
| Domains | 25 |
| Unique sentences | 25,159 (23,390 read by two or more speakers) |
Domains
| Domain | Hours | Share | Clips | Speakers |
|---|
| Literature | 108.2 | 21.6% | 42,082 | 19 |
| Religion | 51.5 | 10.3% | 17,176 | 20 |
| Science and Technology | 45.1 | 9.0% | 13,279 | 19 |
| Culture | 31.6 | 6.3% | 11,920 | 15 |
| Travel and Tourism | 29.4 | 5.9% | 9,138 | 16 |
| Social Media | 25.4 | 5.1% |
Speakers
| Speaker | Gender | Age | Region | Hours | Clips | Domains | Median clip (s) |
|---|
| Speaker-01 | Male | 20-29 | Hyderabad, Sindh | 26.7 | 10,953 | 17 | 8.3 |
| Speaker-02 | Female | 20-29 | South Punjab, Punjab | 25.0 | 13,541 | 13 | 6.0 |
| Speaker-03 | Female | 30-39 | Multan, Punjab | 23.3 | 10,773 | 12 | 7.3 |
| Speaker-04 |
Hours per speaker and domain: speaker_x_domain.csv.
Audio
Measured on every file:
- 169,661 clips: WAV, 44.1 kHz, mono, 16-bit
All clips are 44.1 kHz mono 16-bit PCM WAV.
- Clip length: 0.6 s to 114.5 s, median 9.4 s. 1,692 clips are longer than 30 s.
- Loudness (random sample, 25 clips per speaker): most speakers peak around -1 dBFS with RMS around -16 to -21 dBFS. Speakers 15-19 are quieter (RMS about -31 to -34 dBFS, left as recorded) and speaker 10 is around -30 dBFS. No clipping found.
- Collector-reported recording conditions: background noise about -84 dBFS, SNR about 63 dB, RT60 about 0.18-0.25 s.
Text
- Median 16 words (72 characters) per sentence, up to 108 words.
- 104 sentences contain Latin letters (loanwords, brand names) and 259 contain Latin digits; numbers are not normalised to words.
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
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 and quality checks
- About 60 raw domain labels (including spelling variants) were merged into 25 domains.
- Within each speaker, repeated recordings of the same sentence were reduced to one.
- Every row in
mapping.tsv has its audio file; 75 rows with no audio were removed.
Licence 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
