Sindhi · 52-letter Perso-Arabic script · ~38 million speakers
Sindhi AI training data: speech, text, OCR and annotation
Sindhi has tens of millions of speakers, its own 52-letter alphabet and almost no usable machine-readable data. We collect it across Sindh — Karachi, Hyderabad, Sukkur, Larkana and rural districts — to a written quota, with per-record consent and a training-rights licence.
At a glance
Sindhi training data: the facts a buyer needs first
What a buyer needs to know before scoping a Sindhi speech, text or OCR programme.
| Speakers | ~38 million, overwhelmingly in Sindh province; smaller communities in India (Devanagari script) and the Gulf |
|---|---|
| Script | Extended Perso-Arabic alphabet with 52 letters, written in Naskh rather than Nastaliq — the largest Arabic-script alphabet in use |
| Varieties we field | Vicholi (central Sindh, the written standard), Lari (southern Sindh), Thari (Tharparkar), Utradi and Lasi on request |
| Where we collect | Karachi, Hyderabad, Sukkur, Larkana, Nawabshah, Mirpurkhas and rural Sindh including Thar |
| Capability tier | Deep speech · Deep 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 Sindhi telephony audio from the telephony corpus
- Native Sindhi RLHF preference pairs and toxicity lexicon
- Sindhi rows in the regional speech programme
- Custom: conversational ASR corpora, TTS voices, Sindhi OCR, Sindhi–English parallel text
Built for
- Sindhi ASR for telecom, banking and government IVR in Sindh
- Sindhi OCR for land records, court documents and newspapers
- LLM fine-tuning and evaluation in Sindhi
- Speech translation Sindhi↔Urdu↔English
Why models fail
Why Sindhi breaks tools built for Arabic or Urdu
Sindhi looks like Urdu to a procurement team and behaves nothing like it to a tokenizer, an OCR engine or an acoustic model.
52 letters, many unique
Sindhi adds implosive and aspirated consonants that exist in no other Arabic-script language, each with its own letter. Tokenizers and OCR models built for Arabic or Urdu have never seen these glyphs and split or misread them.
Implosive consonants
Sindhi has four implosive stops that are rare worldwide. Acoustic models transferred from Urdu or Hindi have no phoneme inventory for them and systematically confuse them with plain stops.
Naskh, not Nastaliq
Unlike Urdu, Sindhi is printed in Naskh. That is easier for OCR in principle, but the extended alphabet and heavy diacritics mean Arabic OCR still fails on real Sindhi documents.
Rural distribution
Much of the Sindhi-speaking population is rural and lower-bandwidth. Crowd platforms capture urban Karachi speakers and miss the acoustic and dialect distribution production systems actually see.
What we collect
Sindhi speech, text, image and annotation data
Quota-controlled field collection across Sindh, with transcription and QA by native Sindhi speakers working to a written convention.
Speech for ASR
Spontaneous conversation, telephony, read speech and command utterances across Vicholi, Lari and Thari, balanced by age, gender, urban/rural and device. Verbatim transcripts with a code-switching convention for Urdu and English.
Speech for TTS
Studio-quality single-speaker Vicholi Sindhi voices with phonetically balanced scripts and pronunciation lexicon, 10–40 hours per voice.
Text, SFT and RLHF
Native-authored Sindhi instruction data, dialogues and preference rankings. Sindhi–English and Sindhi–Urdu parallel text across news, government, health, agriculture, finance and everyday domains.
OCR and documents
Printed Sindhi newspapers, books, government forms and handwritten pages with line- and word-level ground truth for the full 52-letter alphabet.
Ready-made
Catalog datasets that include Sindhi
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
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 Sindhi-L1 speakers
ImageNastaliq & Handwriting OCR
Urdu companion set for multi-script document pipelines
SpeechRegional Languages Speech Pack
Saraiki, Balochi, Brahui and Hindko
Specification
Writing a Sindhi spec that survives QA
The decisions that determine whether a Sindhi dataset is usable, made before recording starts.
Choose the variety mix
Vicholi as the standard, with Lari and Thari in stated proportions if the product serves southern Sindh or Tharparkar.
Fix the orthography
Which of the several Sindhi keyboard conventions is canonical; how Urdu and English loanwords are written; whether diacritics are transcribed.
Decide rural share
State a minimum share of rural and low-bandwidth speakers. This is what separates a field corpus from a Karachi crowd sample.
Define acoustic conditions
Telephony versus wideband; outdoor, market and agricultural noise; handset diversity.
Set the acceptance test
Held-out test set from the same quota, WER target, and an inter-annotator agreement floor on double-transcribed audio.
Who buys Sindhi data
Who is buying Sindhi data, and for what
Sindhi demand comes from teams serving Sindh directly and from multilingual programmes that need every major Pakistani language covered to the same standard.
- Government and banking IVR — voice services for a province of 55 million where Sindhi is the working language outside Karachi
- Document digitisation — OCR for land records, courts and newspaper archives in the 52-letter script
- Multilingual ASR programmes — Sindhi as one of the five Pakistani languages in speech-recognition and translation products
- LLM evaluation — native Sindhi test sets and preference data for models that claim South Asian coverage
- Content moderation — Sindhi toxicity and hate-speech data for regional platforms
Answers
Questions buyers ask about Sindhi data
A first QA-passed batch of 50–200 hours of speech or 25k–100k text items in 3–5 weeks from a signed SOW, scaling to a 1,000+ hour programme. Ready-made Sindhi audio is licensable today inside the Pakistani Telephony Corpus.
Vicholi by default, because it is the written standard and the variety of central Sindh. Lari and Thari are fielded to quota when the product serves southern Sindh or Tharparkar.
Not reliably. Sindhi has 52 letters including many that do not exist in Arabic or Urdu, and heavy diacritic use. We supply printed and handwritten Sindhi with line- and word-level ground truth so OCR can be fine-tuned rather than approximated.
Yes. Most of our licensing goes to teams in the US, UK, Europe and Australia. Contracts are in English, invoices in USD by default, and delivery to your own cloud bucket in the region you choose. See Where we deliver.
Yes. Every contributor signs or records a plain-language Sindhi consent granting perpetual, worldwide, sublicensable rights for machine-learning training and derivative-model distribution. Consent artefacts are delivered per record.
Yes. Request a free, unwatermarked Sindhi sample pack with transcripts and the full datasheet, delivered within one business day, no sales call required.
Scoping a Sindhi programme?
Tell us the task, the varieties and the acoustic conditions. We come back with a collection design, a sample pack and an honest view of what data will and will not fix.
Average first response: under 6 business hours. NDAs signed same day.