Regional languages · Saraiki · Balochi · Brahui · Hindko
Saraiki, Balochi, Brahui and Hindko AI training data
Around 40 million people speak these languages, and public AI corpora for them are close to zero. We run standing panels in southern Punjab, Balochistan and Hazara, so a first Saraiki or Balochi batch is weeks away rather than a research project.
At a glance
Saraiki, Balochi, Brahui and Hindko: the facts a buyer needs first
The languages, where they are spoken, and what we can honestly deliver in each.
| Saraiki | ~26 million speakers; southern Punjab — Multan, Bahawalpur, Dera Ghazi Khan, Rahim Yar Khan. Perso-Arabic script with four implosive letters. Deep speech |
|---|---|
| Balochi | ~8 million speakers in Pakistan, more in Iran and Afghanistan; Quetta, Turbat, Gwadar, Kech. Rakhshani and Makrani varieties. Standard speech and text |
| Brahui | ~2.8 million speakers; Kalat, Mastung, Khuzdar. A Dravidian language isolated inside the Iranian and Indo-Aryan zone. Standard speech |
| Hindko | ~5 million speakers; Hazara — Abbottabad, Mansehra — and Peshawar city. Close to Pothwari Punjabi. Standard speech |
| Also on request | Kashmiri and Pahari-Pothwari (AJK), Shina, Balti, Khowar and Burushaski (Gilgit-Baltistan and Chitral) with a scoped 3–6 week lead time |
| Typical pilot | 30–150 hours of speech per language, in 4–6 weeks; creditable against production |
| Licence | Perpetual, worldwide, sublicensable; derivative-model rights stated explicitly. Licensing detail |
| Delivery | Your S3, GCS or Azure bucket with manifests, checksums, QA report and datasheet |
Speaker figures are approximate, drawn from the 2023 census and public sources, and include second-language speakers where noted.
What you can license today
- Saraiki, Balochi, Brahui and Hindko audio in the Regional Languages Speech Pack
- Custom conversational and read-speech corpora per language
- Native-authored text and evaluation sets for LLM coverage claims
- Language-identification training sets across all Pakistani languages
Built for
- Extending Whisper, MMS or in-house ASR to languages with no public audio
- Language-ID and dialect-ID models for South and Central Asia
- Evaluating multilingual LLMs on truly low-resource languages
- Public-sector and NGO voice services in southern Punjab and Balochistan
Why models fail
Why these languages cannot be approximated from Urdu
Each of these is a distinct language, not a dialect of Urdu or Punjabi, and two of them are not even Indo-Aryan.
Saraiki implosives
Saraiki has four implosive consonants, written with letters Urdu does not have. Urdu-trained acoustic models have no category for them, and Urdu keyboards cannot type them, so crowd transcripts silently substitute the wrong letters.
Balochi is Iranian, not Indo-Aryan
Balochi is closer to Kurdish and Persian than to Urdu. Vocabulary, morphology and phonology all differ, and orthography is not fully standardised, so the transcription convention has to be fixed in writing before QA begins.
Brahui is Dravidian
Brahui is a Dravidian language related to Tamil and Telugu, spoken in the middle of Balochistan. Nothing transfers from Urdu beyond loanwords. It is one of the most under-resourced languages with millions of speakers anywhere.
No written internet
These languages are spoken far more than written. There is no web corpus to scrape and no forum archive to mine; the only way to obtain natural data is to record and transcribe it in the field.
What we collect
Regional-language speech, text and annotation data
Field teams in Multan, Bahawalpur, Quetta, Turbat, Kalat and Abbottabad, transcription by native speakers, and QA that checks quota as well as accuracy.
Saraiki speech and text
Spontaneous conversation, read speech and telephony audio across southern Punjab; native-authored Saraiki text and Saraiki–Urdu parallel sentences.
Balochi speech and text
Rakhshani and Makrani speech from Quetta and the Makran coast, with a fixed orthographic convention agreed with the buyer; Balochi–Urdu–English parallel text.
Brahui and Hindko speech
Read and conversational speech to quota in Kalat – Mastung and Hazara respectively, transcribed by native speakers; text on request.
Language-ID and evaluation sets
Balanced multi-language utterance sets across all Pakistani languages for language identification, plus held-out native test sets for LLM and ASR evaluation.
Ready-made
Catalog datasets that include these languages
Each ships with a datasheet, per-record consent artefacts and a free sample pack. Subsets by language are available on request.
Regional Languages Speech Pack
Saraiki, Balochi, Brahui and Hindko
SpeechPakistani Telephony Corpus
8 kHz call-center audio in Urdu, Punjabi, Sindhi, Pashto
SpeechPakistani-Accented English
L1-tagged, including Saraiki and Pashto L1 speakers
TextRLHF Preference Pairs
Native rater rankings in Urdu, Punjabi, Sindhi, Pashto
SpeechWake Word & Command Set
Urdu, Punjabi, Pashto and English commands
TextUrdu–English Parallel Corpus
Human-translated, six domains
Specification
Writing a regional language spec that survives QA
With languages this under-resourced, the specification is where the budget is won or lost.
Pick languages and proportions
One language deeply, or several to a common protocol for a language-ID or coverage programme. State hours per language.
Agree the orthography first
Especially for Balochi and Brahui, where spelling is not standardised. We propose a convention, you approve it, and QA enforces it.
Decide the register
Read speech from prompts is fast but unnatural; conversational speech is what production hears. Most programmes need a stated mix.
Define acoustic conditions and devices
Rural low-bandwidth handsets are the norm here, not the exception. State the share you want.
Set the acceptance test
Held-out native test set, WER or agreement target, and a re-collection clause for out-of-quota material.
Who buys regional language data
Who needs Saraiki, Balochi and Brahui data
Demand comes from three directions: products serving these regions, multilingual models that claim coverage, and research programmes that need real low-resource benchmarks.
- Multilingual ASR and LLM coverage — frontier labs and speech companies extending language lists past the top hundred
- Language identification — telecom, moderation and search systems that must route Pakistani speech to the right model
- Public services and NGOs — health, agriculture and disaster-response voice services in southern Punjab and Balochistan
- Academic and benchmark work — held-out native test sets for languages with no existing evaluation data
- Cross-border Balochi — Pakistani Balochi as the base for systems serving Iranian and Afghan Balochi speakers
Answers
Questions buyers ask about regional language data
Yes, with honest lead times. Standing panels in Quetta and Kalat let us deliver 30–150 hours per language within 4–6 weeks. Larger programmes are scoped in phases so quota and orthography are validated on the first batch before headcount scales.
Almost none of usable size, and what exists is read speech or religious text. For production systems the practical route is commissioned collection; see Free vs commissioned datasets for how to decide.
We propose a written convention based on the most widely used community and academic practice, you approve it, and every transcriber works to it. Inter-annotator agreement on double-transcribed samples is reported in the datasheet.
Yes. Most of our licensing is to teams outside Pakistan. English-language contracts, USD invoicing by default, and delivery to your cloud region of choice. See Where we deliver for compliance detail.
Consent is read aloud in the contributor's own language and recorded, with written confirmation where possible. Every record carries the consent version and a withdrawal route. The method is described on Consent & Fair Pay.
On request. These are bespoke recruitments in Gilgit-Baltistan and Chitral with a 3–6 week lead time and a scoping conversation first, so that the quote is honest.
Which language is missing from your model?
Name it. If it is spoken in Pakistan we will tell you what it takes to collect it, how long it will take, and what it will cost — before you commit to anything.
Average first response: under 6 business hours. NDAs signed same day.