Pashto · Northern (Yusufzai) and Southern (Kandahari) · ~40 million speakers in Pakistan, ~60 million with Afghanistan
Pashto AI training data: dialect-labelled speech and text
Pashto spans Pakistan and Afghanistan, splits into two dialect groups that differ systematically in pronunciation, and has no enforced spelling standard. Corpora that ignore either fact fail in production. We collect Pashto in Peshawar, Mardan, Swat and the border belt, labelled by dialect group, transcribed to a written orthographic convention, consented per record.
At a glance
Pashto training data: the facts a buyer needs first
The facts that decide whether a Pashto corpus will work for your users.
| Speakers | ~40 million in Pakistan (Khyber Pakhtunkhwa, Balochistan, Karachi); ~20 million more in Afghanistan, where it is an official language; diaspora in the UAE, UK, Germany and North America |
|---|---|
| Script | Perso-Arabic with extra letters for retroflex and other Pashto-specific sounds; spelling of common words varies between writers and regions, so transcription needs an enforced convention |
| Dialect groups we field | Northern / Yusufzai (Peshawar, Mardan, Swat, Charsadda) — Deep. Southern / Kandahari (Quetta, Waziristan, border belt) — Standard. Central varieties on request |
| Where we collect | Peshawar, Mardan, Swat, Charsadda, Nowshera, Kohat, Bannu, Quetta and Pashtun neighbourhoods of Karachi |
| Capability tier | Deep speech · Standard text · Standard image/video |
| Typical pilot | 50–200 hours of speech, or 25k–100k text items, in 3–5 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
- 8 kHz Pashto telephony audio in the telephony corpus
- Pashto wake-word and command utterances
- Native Pashto RLHF preference pairs and severity-graded toxicity lexicon
- Pashto-L1 speakers in the accented-English set
- Custom: dialect-balanced conversational ASR corpora, TTS voices, Pashto–English parallel text
Built for
- Pashto ASR for telecom, banking and humanitarian voice services in Pakistan and Afghanistan
- Language and dialect identification across Pashto, Dari and Urdu
- LLM coverage, evaluation and safety for Pashto
- Speech translation Pashto↔English for interpreting and media monitoring
Why models fail
Why Pashto corpora fail: dialect and orthography
Two problems account for most Pashto model failures, and both are decided at specification time.
The dialect problem comes first
Northern (Yusufzai) and Southern (Kandahari) Pashto differ systematically in consonants — the same word is pronounced with different sibilants and fricatives. A model trained on one group degrades sharply on the other. Every record we deliver carries a dialect-group label and a country attribution.
Orthography quietly ruins corpora
Pashto spelling is not enforced: the same word appears in several forms across writers and regions. Without a written convention, transcribers fragment the vocabulary, tokenizers see three words where there is one, and WER is inflated by spelling, not by acoustics. We fix the convention with you before QA starts.
Open Pashto data is the wrong distribution
Public Pashto corpora skew formal, male, urban and read-from-script — the opposite of what call centres, health lines and voice assistants hear. Female speakers, rural districts and spontaneous speech require deliberate, community-based recruitment.
Cross-border transfer works, if labelled
Pashto collected in Pakistan transfers directly to Afghan Pashto in the same dialect group and closely to Dari-influenced varieties. That is only useful if dialect and country are labelled per record so evaluation splits can be built per group.
What we collect
Pashto speech, text, image and annotation data
Field teams in Khyber Pakhtunkhwa and Quetta, female recruiters for female contributors, transcription by native speakers to an agreed orthography, and QA that checks dialect quota as well as accuracy.
Speech for ASR
Spontaneous conversation, telephony, read speech and commands, quota-balanced across dialect group, age, gender, urban/rural and device. Verbatim transcripts in the agreed orthography with Urdu/English code-switch tagging and dialect label per record.
Speech for TTS
Studio Northern-Pashto voices with phonetically balanced scripts covering retroflex and dialect-specific phonemes, pronunciation lexicon, 10–40 hours per voice.
Text, SFT and RLHF
Native-authored Pashto instruction data and preference rankings; Pashto–English and Pashto–Urdu parallel text in health, humanitarian, government, news and everyday domains; toxicity and safety lexicon.
Language-ID and evaluation sets
Balanced Pashto / Dari / Urdu / Hindko utterance sets for language and dialect identification, and held-out native test sets per dialect group.
Ready-made
Catalog datasets that include Pashto
Each ships with a datasheet, per-record consent artefacts and a free sample pack. Subsets by language are available on request.
Pakistani Telephony Corpus
8 kHz call-center audio in Urdu, Punjabi, Sindhi, Pashto
SpeechWake Word & Command Set
Urdu, Punjabi, Pashto and English commands
TextRLHF Preference Pairs
Native rater rankings in Urdu, Punjabi, Sindhi, Pashto
TextSafety & Toxicity Lexicon
Severity-graded terms in Urdu, Punjabi, Pashto, Sindhi
SpeechPakistani-Accented English
L1-tagged, including Pashto-L1 speakers
SpeechRegional Languages Speech Pack
Hindko and neighbouring languages of KP
Specification
Writing a Pashto spec that survives QA
A usable Pashto specification states five things. Missing any of them is how a corpus ends up unusable.
Dialect groups and proportions
Northern only, or Northern plus Southern in stated shares. Label every record. Build evaluation splits per group.
Country attribution
Pakistan-collected, and whether Afghan-origin speakers resident in Pakistan are in or out of quota.
Orthographic convention
The spelling standard transcribers follow, including Urdu and English loanwords and Pashto-specific letters. Agreed in writing before QA.
Speaker quotas
Age, gender (state a minimum female share and how it will be achieved), urban/rural, device.
Acceptance test
Held-out native set per dialect group, WER target, inter-annotator agreement floor on double-transcribed audio.
Who buys Pashto data
Who buys Pashto data
Demand comes from products serving Pakistan and Afghanistan, humanitarian and public-sector voice services, and multilingual models that need real Pashto coverage rather than a claim.
- Telecom and banking IVR — Pashto voice services for Khyber Pakhtunkhwa and Balochistan, and for Afghan markets
- Humanitarian and health lines — NGO and government helplines where Pashto is the only language callers speak
- Media monitoring and translation — Pashto ASR and translation for broadcast and open-source monitoring
- LLM coverage and evaluation — native Pashto test sets and preference data for South and Central Asian coverage
- Diaspora services — Pashto-speaking communities in the UAE, UK, Germany and North America
Answers
Questions buyers ask about Pashto data
Match your users. Northern (Yusufzai) covers Peshawar, Mardan and Swat and most of Pakistan's Pashto speakers; Southern (Kandahari) covers Quetta, Waziristan and southern Afghanistan. Products serving both need both, in stated proportions, with a dialect label on every record.
Within a dialect group, speech transfers directly across the border. Vocabulary and register differ somewhat, and country attribution matters for compliance and for evaluation, so we label country and dialect group per record rather than treating Pashto as one pool.
We propose an orthographic convention based on the most widely used practice, you approve it, and every transcriber works to it. Inter-annotator agreement on double-transcribed samples is reported in the datasheet so you can see the convention holding.
Yes, with female recruiters and field staff, community-based recruitment, and recording arrangements that respect local norms. State the minimum female share you need in the spec; we will tell you honestly what timeline it implies.
A first QA-passed batch of 50–200 hours in 3–5 weeks from Northern-Pashto panels; Southern Pashto ramps over 2–3 weeks. Programmes of 1,000+ hours are phased so dialect quota and orthography are validated on the first batch.
Yes. Most of our licensing goes to teams in the US, UK, Europe and Australia. English contracts, USD invoicing by default, delivery to your cloud region with a DPA. See Where we deliver.
Which Pashto is your model failing on?
Tell us the dialect group, the task and the acoustic conditions. We come back with a collection design, dialect-labelled samples and an honest timeline.
Average first response: under 6 business hours. NDAs signed same day.