Persian (Farsi) AI Training Data Services
Farsi-speaking contributors record, transcribe, label and rate your data under fixed character rules, with calibration before production and gold tasks during it.
- Languages
- 19
- Languages
- Vetted Native Linguists
- 700+
- Vetted Native Linguists
- Localization Tasks / Year
- 100K+
- Localization Tasks / Year
- Founded
- 2016
- Founded
El Turco provides Persian (Farsi) AI training data for enterprise AI teams and technology companies, covering speech collection, transcription, text annotation, RLHF, content moderation, image labeling and custom MT and LLM datasets. Every dataset is collected, labeled and checked by native Farsi speakers. Our Istanbul production team runs each project, with headquarters in London and client data kept under ISO/IEC 27001.
Persian AI training data is the Farsi speech, text and image material, with its labels, that teams use to train, fine-tune and evaluate models. Persian web text is noisy. Arabic and Persian characters get mixed, the half space inside words goes missing and many users type Finglish, Persian written in Latin letters. Enterprises that build or adapt their own models are our main clients, and LSPs use the same Farsi annotators as a white-label data partner.
Shared workflow: AI Data Solutions.
Scope
What does Persian AI training data cover?
Speech data collection
Scripted and spontaneous Persian recordings from native speakers across ages, genders and acoustic environments, covering formal read speech and colloquial Tehran-style speech. Files ship as WAV or FLAC between 16 and 48 kHz, each with speaker and environment metadata.
Transcription
Time-aligned Persian transcripts in TextGrid, SRT or VTT, written to agreed conventions for colloquial forms, fillers, digits and English loanwords. Word error rate spot-checks measure each batch against the acceptance criteria you set.
Text annotation and NER
Entity, intent, sentiment and topic labels on Persian sentences and documents, exported in NFC Unicode as JSONL, CSV/TSV or a schema you define. Image labeling covers Persian signage, forms and product packaging.
RLHF and LLM evaluation
Farsi-speaking raters compare model outputs, grade how natural and accurate each answer is and mark replies that sound translated from English or drift into Arabic vocabulary.
Content moderation data
Labels for hate speech, toxicity, misinformation and sentiment in Persian posts, comments and chats, covering slang, Finglish and Persian-English code-switching.
Parallel and LLM training data
Parallel corpora, domain datasets and evaluation sets in both directions between English and Persian, translated under ISO 17100 and post-edited under ISO 18587 where MT output is used.
Persian specifics
What makes Persian different for AI training data?
Persian versus Arabic characters
Persian and Arabic encode yeh and kaf as different characters even though the letters look nearly identical. Scraped text mixes them, which splits one word into two tokens and breaks deduplication. We fix a normalization rule before production and deliver every dataset with Persian characters only.
The half space
Persian uses the zero-width non-joiner inside words, as in the verb prefix mi- and the plural suffix -ha. Annotators who drop it or type a full space create spelling variants of the same word. Guidelines state where the half space belongs and QA checks for it in every batch.
Written versus spoken Persian
Spoken Persian shortens many verb endings and word forms, so it often looks quite different from the written standard on the page. Transcription guidelines say whether to write speech as spoken or in standard spelling, so ASR data stays consistent.
Finglish and mixed scripts
Many users type Persian in Latin letters, with no fixed spelling. Moderation and sentiment sets need these samples because real users write this way. Annotators label the text as written, with an optional Persian-script field when the client wants one.
Right-to-left text in labeling tools
Persian runs right to left while digits, URLs and English terms inside it run left to right. Entity spans in bidirectional lines are easy to misplace. We test the labeling tool with Persian samples in calibration and check span offsets in QA.
Variety and register
Iranian Persian, Dari and Tajik are separate written standards. Tajik uses Cyrillic. Business answers use formal "shoma". The target variety and form of address are fixed per project and raters mark every slip.
Process
How we run a Persian AI training data project
Turnaround is agreed at scoping, with a 24-hour option for small urgent updates.
Step 1
Scoping and specification
Scoping covers your model goals, data volume, formats and acceptance criteria. For speech we agree the speaker mix, environments and the written or spoken transcription style. For text we agree the label set, output schema and normalization rules.
Step 2
Guidelines and calibration
Your team and ours draft the annotation guidelines together and give each batch a version number. Calibration rounds settle Persian edge cases such as half spaces, Finglish and entity spans in bidirectional text before production starts.
Step 3
Collection and annotation
Every Persian annotator and speaker is a vetted native contributor working under NDA. No recording starts until the speaker has given consent.
Step 4
Layered data QA
Hidden gold tasks run through every batch and show how each annotator performs. Each delivery includes an inter-annotator agreement report. Parallel data also passes What's Wrong?, our own QA tool: rules check terms, numbers and Persian characters, an AI judge re-reads risky segments and a human signs off.
Step 5
Secure delivery
Before delivery a reviewer checks each batch and formats it for your pipeline. Batches go out by secure transfer, API or directly into your platform. The data of every client is processed in its own environment under ISO/IEC 27001.
Industries
Who buys Persian ai training data?
Industry
Technology & IT
AI labs and product teams at SaaS and enterprise software companies building or testing Persian speech recognition, chat and search.
Industry
Gaming & Esports
Gaming and community platforms moderating Persian player chat, where Finglish and slang are common.
Industry
E-Commerce & Retail
Online marketplaces and retailers that label Persian product listings, reviews and customer messages to train search and support models.
Related
Language page
Persian translation and localization
Native Persian (Farsi) linguists for right-to-left software, game, e-commerce and marketing content, plus Persian AI training data.
Same language
Persian Game Localization
Persian in-game text, right-to-left UI, store pages and in-build LQA for mobile, PC and console titles, in formal or colloquial Farsi.
Same language
Persian E-Commerce Localization
Persian product catalogs, right-to-left storefronts, checkout and campaigns by native Farsi linguists, with MTPE for large catalogs.
Same language
Persian Software Localization
Persian UI strings, right-to-left review, half-space and character QA, help content and release notes for SaaS and enterprise software.
Same service
Turkish AI Training Data
Turkish speech collection, text annotation, RLHF, moderation and MT training data from native speakers, with gold tasks and agreement reporting.
Same service
Arabic AI Training Data
Arabic speech collection, transcription, text annotation, RLHF and MT corpora across MSA and the major dialects, labeled by native speakers.
FAQ
Persian ai training data FAQ
Which Persian AI training data services does El Turco offer?
El Turco builds Persian datasets for speech recognition, NLP and LLM work: recordings, transcripts, text labels, RLHF rankings, moderation labels, image annotations and parallel MT and LLM data. Native Farsi speakers produce every dataset. Gold tasks, inter-annotator agreement and word error rate spot-checks control quality. Each client has its own processing environment, and all Persian data is handled under ISO/IEC 27001.
How does El Turco handle mixed Persian and Arabic characters?
El Turco fixes a normalization rule for Persian and Arabic characters during scoping. Arabic yeh and kaf are converted to their Persian forms, half spaces are kept where Persian spelling needs them and text is delivered in NFC Unicode. When source data or a client pipeline reintroduces Arabic characters, El Turco reports the affected samples so engineers can fix the step before training.
What delivery formats does El Turco use for Persian data?
El Turco ships Persian text data in NFC Unicode as JSONL, CSV/TSV, XLIFF or a schema you define. Speech files are WAV or FLAC between 16 and 48 kHz with speaker and environment metadata, and time-aligned transcripts use TextGrid, SRT or VTT. RLHF rankings and ratings follow the format of your pipeline. Data reaches you by secure transfer, API or your own platform.
Do Persian moderation datasets include Finglish and slang?
Persian moderation and sentiment datasets from El Turco include informal text as users write it: slang, colloquial spelling, Finglish typed in Latin letters and Persian-English code-switching. Native Farsi annotators label each sample as written and take its cultural context into account. A Persian-script field can be added when your model needs both versions of each sample.
Can LSPs and data vendors subcontract Persian AI data work to El Turco?
LSPs and data vendors can subcontract Persian speech, annotation and RLHF work to El Turco and deliver it under their own brand. Vendor managers receive the same Farsi annotators, calibration rounds and quality reports as enterprise clients, inside the platform and workflow of the LSP. A free 300-word test translation is available for parallel data work. Pricing is per-word, hourly or project-based, quoted after scoping.
How does a Persian AI training data project begin?
A Persian AI training data project begins with the contact form, where you describe the data type, volume, formats and acceptance criteria. El Turco then runs a free pilot so your team can measure quality with its own metrics. Guidelines and calibration rounds are completed before full production. Turnaround is agreed at scoping, with a 24-hour option for small urgent updates.
Ready to Scope Persian AI Training Data?
Share the content type, volume and timeline for a scoped plan. A free pilot comes first if you want proof before committing. Pricing is per-word, hourly or project-based, quoted after scoping.