Indian language careers: teaching, media, content and AI language data

Indian Languages and Literature (Hindi / Sanskrit / Urdu / Malayalam / Tamil / Telugu / Kannada / Bengali / Marathi ...) (BA) · Official language, exams and careers · free study notes for exams, placements and interviews

In short

A BA in an Indian language leads to four broad employment markets: teaching and academia; news media, broadcasting and publishing; commercial content, translation and localisation; and the newest and fastest-growing one, AI language data work (transcription, annotation, machine-translation post-editing, prompt and preference data for Indic LLMs). Demand is driven by the National Education Policy 2020's push for mother-tongue instruction, by OTT and short-video platforms dubbing into 10-15 Indian languages, and by Bhashini, AI4Bharat and private Indic-LLM builders who need labelled data in all 22 Eighth Schedule languages. University papers on "language and employment" or "applied/functional language" examine these career structures, employer types, rate and contract practice and the Acts that govern publishing and broadcasting. Interviews at localisation firms, publishers and data-services companies test both language skill and the measurable side — per-word rates, throughput, subtitle reading speed, word error rate and inter-annotator agreement.

Explained

The four career tracks and their entry gates

Teaching: school (TGT/PGT routes, state and central recruitment), college (UGC NET/SET plus PhD for permanent posts), coaching and ed-tech. Covered in detail by the separate notes on teacher eligibility tests and UGC NET.
Government language work: Hindi Translator/Hindi Officer and Rajbhasha cells — a distinct track with its own note; Sanskrit/Persian/Urdu scholars also go into ASI, archives, manuscript libraries and the National Mission for Manuscripts.
Media and publishing: reporter, sub-editor, copy editor, proof reader, news anchor, radio jockey, script and dialogue writer, dubbing director, publisher's editor.
Commercial content and localisation: translator, transcreator, UX/microcopy writer, SEO content writer, technical writer, localisation project manager, LQA (language quality assurance) reviewer.
AI language data: transcriber, annotator, MT post-editor, linguistic rule writer, prompt writer, RLHF preference rater, red-teamer, computational-linguistics data lead.
Eligibility is mostly skill-tested, not degree-gated, outside teaching and government: employers run a translation test, a 300-word writing test or a paid pilot batch.

Media, broadcasting and publishing

Print: large language dailies (Dainik Jagran, Amar Ujala, Malayala Manorama, Mathrubhumi, Eenadu, Sakshi, Daily Thanthi (Dina Thanthi), Lokmat, Anandabazar Patrika, Vijaya Karnataka, Rajasthan Patrika) hire trainee sub-editors; the work is headline writing, translation of agency copy (PTI/IANS/Bhasha), page proofing and style-guide discipline.
Broadcast: Prasar Bharati (AIR/Doordarshan) empanels casual news readers, translators, announcers and compères by voice test; private news channels hire anchors, input/output desk and ticker writers.
Regulation to know: registration of newspapers and periodicals is now under the Press and Registration of Periodicals Act, 2023 (which replaced the Press and Registration of Books Act, 1867) with the Press Registrar General of India; digital news publishers follow the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 — Code of Ethics, grievance officer, three-tier redressal; films are certified by the CBFC under the Cinematograph Act, 1952 (amended 2023, with age-based UA 7+/13+/16+ categories); the Press Council of India handles print complaints.
Publishing: acquisitions and commissioning, copy editing, translation lists. Institutional employers: Sahitya Akademi, National Book Trust, Central Institute of Indian Languages (CIIL) Mysuru and its National Translation Mission, state Sahitya/Granth Akademis, NCERT and SCERT textbook units; trade houses include Rajkamal, Vani, DC Books, Mathrubhumi Books, Manjul, HarperCollins and Penguin's Indian-language imprints.
Awards that function as professional credentials: Sahitya Akademi Award (₹1,00,000), its Translation Prize (₹50,000), Bhasha Samman, Yuva Puraskar; Jnanpith (₹11 lakh). Amounts are revised from time to time.
Copyright basics every editor is asked: the Copyright Act, 1957 — translation is a separate exclusive right of the author; literary-work term is the author's life + 60 years; moral rights (paternity and integrity) survive assignment; a compulsory licence to publish a translation can be sought after a statutory waiting period, shorter for Indian languages; assignment must be in writing and specify the work, rights, term and territory.

Translation, localisation and content services

Translation renders meaning; transcreation rewrites for cultural effect (ads, slogans, app onboarding); localisation (L10n) adapts the whole product — dates, numerals, currency, honorifics, name order, sort order, text expansion.
Workflow: source analysis → TM/glossary setup → translation or MT → post-editing → LQA review → DTP/engineering → client sign-off. CAT tools (Trados Studio, memoQ, Smartcat, MateCat, Phrase) store a translation memory (TM) and termbase, bill by new/fuzzy/repeat word bands, and are the single most asked tool skill.
MTPE (machine-translation post-editing): *light* post-editing fixes only meaning-breaking errors; *full* post-editing produces publication quality. Paid at 50-70% of the from-scratch per-word rate because throughput roughly doubles.
Subtitling: spotting (timing) + translation into SRT/VTT files; limits are typically 2 lines, about 42 characters per line, 1-7 seconds per cue, and a reading speed around 17-20 characters per second (platform style guides, e.g. Netflix's per-language Indic guides, fix the exact figures). Tools: Subtitle Edit, Aegisub, Ooona.
Dubbing and voice: adaptation (lip-sync/phonetic matching), dialogue writing, voice direction, studio recording; billed by *loop*, reel, hour or finished minute. Hubs: Mumbai (Hindi), Chennai, Hyderabad, Kochi, Kolkata.
Employers: Reverie Language Technologies, Process9, Devnagri, Crystal Hues, Braahmam, Lionbridge, RWS, TransPerfect, Welocalize, plus in-house teams at banks, insurers, e-commerce, government portals and OTT platforms (JioHotstar, Netflix, Prime Video, ZEE5, SonyLIV, Aha, Sun NXT).
Script and encoding practice: use Unicode/UTF-8 (Devanagari U+0900-U+097F, Tamil from U+0B80, and so on); ISCII (IS 13194) and legacy font-based "Kruti Dev/Shree Lipi" encodings are legacy and must be converted, not typed fresh. Input: InScript layout, Google Input Tools, phonetic/transliteration keyboards, Keyman. Know ISO 15919 transliteration and IAST for Sanskrit.

AI language data work — the new mass employer

Why it exists: Indic languages are low-resource. Models need parallel corpora, speech with transcripts, instruction and preference data, and safety data in each language and dialect — none of which can be scraped at the quality needed.
Public ecosystem: Bhashini (Digital India Bhashini Division, MeitY) with ULCA as its open datasets/API repository and *Bhasha Daan* crowdsourcing (Suno/Bolo/Likho/Dekho India); AI4Bharat at IIT Madras (IndicCorp, IndicTrans2 across 22 scheduled languages, IndicVoices, Airavata); Project Vaani (IISc with Google) for district-wise speech. Private: Sarvam AI, Krutrim, CoRover's BharatGPT, Tech Mahindra's Project Indus, plus data vendors Karya, Indika AI, iMerit, Shaip, DesiCrew.
Task types: (1) speech data collection — scripted read speech, spontaneous prompts, dialect coverage; (2) transcription with convention sheets (disfluencies, code-mixing, numerals, tags); (3) text annotation — POS tagging, NER, dependency/treebank work, sentiment, intent, toxicity; (4) parallel-data creation and translation QA; (5) prompt/response writing and RLHF preference ranking; (6) red-teaming for unsafe output in the target language; (7) romanisation/transliteration and code-mixed (Hinglish) normalisation.
Quality control: written guidelines, gold/honeypot items, double annotation with inter-annotator agreement (Cohen's/Fleiss' kappa), adjudication by a senior linguist, and acceptance thresholds in the purchase order.
Evaluation metrics a language graduate must be able to read: WER/CER for speech, BLEU, chrF, TER, COMET for translation, and human adequacy/fluency scores on a 1-5 scale. BLEU is a surface n-gram overlap score and is weak for morphologically rich Indian languages — chrF and COMET or human scoring are preferred.
Pay structure: usually per audio hour, per 1,000 words, per task or per validated unit, with a quality multiplier — so throughput and rejection rate decide actual earnings more than the headline rate.

Policy and institutional drivers to quote in exams

NEP 2020: mother tongue/regional language as medium in early grades, three-language formula, Indian Knowledge Systems, and translation of higher-education material — the ASMITA project (UGC with the Bharatiya Bhasha Samiti) targets thousands of textbooks across the 22 scheduled languages. It expands demand for teachers, translators and textbook editors simultaneously.
Classical language status now covers 11 languages after Marathi, Pali, Prakrit, Assamese and Bengali were added in October 2024 (with Tamil, Sanskrit, Telugu, Kannada, Malayalam, Odia) — this carries research centres, chairs and fellowship funding.
Eighth Schedule: 22 languages — the standard denominator for government language technology and publishing programmes. Official-language policy itself is a separate topic.
Employment law for this workforce: the four Labour Codes are in force from 21 November 2025; the Code on Social Security, 2020 brings gig and platform workers (which covers much app-based transcription and annotation work) into a registration and aggregator-contribution framework, and a written appointment letter is now required — relevant because most language work is freelance or fixed-term.

Money, invoicing and tax for freelance language work

Rate bases: per source word (translation), per target word, per audio/video minute (subtitling, transcription), per hour (voice, consulting), per page of 250 words (publishing copy editing), per loop (dubbing), per task (annotation). Always confirm *source or target* word count — Indian-language targets usually expand.
Indicative 2026 market rates — they vary widely by language pair, domain and client (direct client versus agency): Indian-language translation roughly ₹1.5-₹5 per word, MTPE ₹0.8-₹2.5, subtitling ₹40-₹150 per video minute, transcription ₹20-₹60 per audio minute, annotation ₹150-₹400 per working hour, voice recording ₹1,500-₹6,000 per studio hour. Treat these as negotiating ranges, not published tariffs.
GST: services attract 18% (the slabs since 22 September 2025 are 5%/18%/40%); registration threshold for services is ₹20 lakh turnover (₹10 lakh in special-category states); export of services is zero-rated — bill without GST under a Letter of Undertaking and keep FIRC/BRC proof.
Income tax: the Income-tax Act, 2025 applies from 1 April 2026. Freelance language work is professional income; the presumptive scheme for specified professions deems 50% of gross receipts as profit up to ₹50 lakh (₹75 lakh where cash receipts are not more than 5%). Clients deduct TDS on professional fees at 10% (claim it against your liability); advance tax is payable in four instalments.
Contract clauses to check: scope and word count, rate and payment term (30/45/60 days), rejection and rework policy, confidentiality, copyright assignment versus licence, credit line (important for literary translation royalties), and late-payment interest.

Key points

Four markets - teaching, media/publishing, commercial content/localisation, AI language data; only teaching and government language work are degree-gated, the rest are skill-tested.
Transcreation - cultural rewriting for effect, priced far above plain translation.
MTPE - post-editing machine output; paid at about 50-70% of full translation rate for roughly double throughput.
CAT tool - Trados/memoQ/Smartcat; stores translation memory and termbase and drives fuzzy-match pricing.
TM match bands - new, fuzzy (75-99%), 100% and repetitions are each paid at a different percentage of the base rate.
Subtitle limits - about 42 characters per line, 2 lines, 1-7 s per cue, ~17-20 characters per second reading speed.
Unicode/UTF-8 - the only correct encoding; Kruti Dev-type font hacks and ISCII are legacy and need conversion.
ISO 15919 / IAST - standard romanisation schemes for Indic scripts.
Bhashini + ULCA - MeitY's translation mission and its open dataset/API repository; Bhasha Daan crowdsources data.
AI4Bharat - IIT Madras group behind IndicCorp, IndicTrans2 and IndicVoices for the 22 scheduled languages.
WER - (substitutions + deletions + insertions) / reference words; the headline ASR quality number.
Kappa - chance-corrected inter-annotator agreement; ≥0.61 is usually read as substantial.
BLEU's weakness - n-gram overlap penalises morphologically rich Indian languages; prefer chrF/COMET or human adequacy-fluency.
Copyright Act, 1957 - translation is a separate right; literary term life + 60 years; moral rights survive assignment.
PRP Act, 2023 - replaced the 1867 PRB Act for periodical registration; IT Rules 2021 Code of Ethics binds digital news publishers.
NEP 2020 and ASMITA - mother-tongue instruction and mass textbook translation into the 22 scheduled languages.
11 classical languages - after Marathi, Pali, Prakrit, Assamese and Bengali were added in 2024.
Tax and GST - 18% GST on services (₹20 lakh services threshold), exports zero-rated, 10% TDS on professional fees, 50% presumptive profit for professions under the Income-tax Act, 2025.

Formulas and examples

Fee = word count × rate per word (₹). Subtitling fee = video minutes × rate per minute.
Effective hourly rate = total fee ÷ hours actually spent (₹/h) - the only honest comparison between jobs.
Daily output needed = target monthly income ÷ (rate per word × working days) (words/day)
Freelance rate floor = (annual income target + annual business costs) ÷ billable hours per year (₹/h)
CPS (reading speed) = characters in the cue ÷ cue duration in seconds; keep ≤ ~17-20 CPS
WER = (S + D + I) ÷ N, where S = substitutions, D = deletions, I = insertions, N = words in the reference; CER is the same on characters
Cohen's kappa κ = (Po − Pe) ÷ (1 − Pe); Po = observed agreement, Pe = agreement expected by chance
Invoice total = fee × 1.18 (GST 18%); cash received = fee − 0.10 × fee (TDS) + GST collected
Presumptive professional profit = 0.50 × gross receipts (specified professions, within the turnover limit)
Unit cost of speech data = collection cost + (transcription hours per audio hour × transcriber rate) + QA cost, all per audio hour

Example 1 - translation invoice, GST and TDS. 12,000 source words of a Hindi user manual at ₹2.50/word.

Fee = 12,000 × 2.50 = ₹30,000. GST at 18% = ₹5,400 → invoice ₹35,400.
Client deducts TDS at 10% on the fee (not on GST) = ₹3,000 → cash received = 35,400 − 3,000 = ₹32,400.
Of that, ₹5,400 is GST to be remitted, so retained = ₹27,000, plus a ₹3,000 TDS credit to set off against tax.
If 3,000 of those words were 100% TM matches paid at 25%, fee = (9,000 × 2.50) + (3,000 × 0.625) = 22,500 + 1,875 = ₹24,375.

Example 2 - subtitling throughput and reading speed. A 48-minute episode at ₹60 per video minute, 620 cues.

Fee = 48 × 60 = ₹2,880. Working at 5 minutes of content per hour → 48/5 = 9.6 hours → effective rate = 2,880/9.6 = ₹300/hour.
A cue of 58 characters lasting 3.2 s gives CPS = 58/3.2 = 18.1 - above a 17 CPS cap, so the line must be condensed to 17 × 3.2 ≈ 54 characters or the cue extended.

Example 3 - word error rate. A Marathi ASR output against a 120-word reference shows 7 substitutions, 3 deletions, 2 insertions.

WER = (7 + 3 + 2)/120 = 12/120 = 0.10 = 10%; word accuracy = 90%.
After the vendor's correction pass only 4 errors remain → WER = 4/120 = 3.33%, i.e. a 66.7% relative reduction.

Example 4 - inter-annotator agreement. Two annotators label 200 Tamil sentences as positive/negative; they agree on 170. Annotator A marked 120 positive, B marked 130 positive.

Po = 170/200 = 0.85.
Pe = (120/200)(130/200) + (80/200)(70/200) = 0.60×0.65 + 0.40×0.35 = 0.39 + 0.14 = 0.53.
κ = (0.85 − 0.53)/(1 − 0.53) = 0.32/0.47 = 0.68 - substantial agreement; usable, but the guideline should be tightened before scaling.

Example 5 - costing a speech-data project and a freelancer's rate floor. 500 hours of Bhojpuri speech; collection ₹250 per validated audio hour; transcription takes 6 hours of work per audio hour at ₹120/hour; QA ₹100 per audio hour.

Per audio hour = 250 + (6 × 120) + 100 = 250 + 720 + 100 = ₹1,070 → 500 hours = ₹5,35,000.
A freelancer wanting ₹6,00,000 a year with ₹60,000 of costs and 1,200 billable hours needs (6,00,000 + 60,000)/1,200 = ₹550/hour; at 350 words/hour that is ₹1.57/word as the floor.

Track	Typical entry gate	Pay basis	Growth path
School/college teaching	TET/CTET, UGC NET + PhD	Monthly pay level (e.g. TGT level 7, PGT level 8 in central schools)	HOD, principal, professor
Print/broadcast media	Writing and translation test	Monthly salary, stringer per-story	Sub-editor → desk head → editor
Translation/localisation	Paid test batch	Per word, per minute, per hour	Reviewer → LQA lead → project manager
AI language data	Guideline test, gold items	Per task/audio hour + quality bonus	Senior annotator → linguist → data lead

	Human translation	Full MTPE	Light MTPE
Throughput	~300-400 words/h	~700-1,000 words/h	~1,200+ words/h
Rate	100% of base	~60-70%	~50%
Output quality	Publication	Publication	Comprehensible only
Best for	Literature, ads, legal	Manuals, support content	Internal, high-volume text

Where it is used

OTT and broadcast localisation: subtitle and dub tracks in 8-15 Indian languages for JioHotstar, Netflix, Prime Video, ZEE5, SonyLIV, Aha and Sun NXT; roles in spotting, adaptation, voice direction and LQA.
Government language technology: Bhashini/ULCA datasets, CIIL's National Translation Mission, NCERT/SCERT textbook translation, state e-governance portals and court-judgment translation projects.
Banking, insurance and telecom: regulator-driven multilingual disclosure, IVR scripts, app microcopy and chatbot utterance sets in regional languages.
AI data vendors and labs: Karya, Indika AI, iMerit, Shaip, DesiCrew supplying Sarvam AI, Krutrim, BharatGPT, Project Indus and global labs with Indic speech, text and preference data.
Publishing and awards economy: Rajkamal, DC Books, Mathrubhumi, NBT and Sahitya Akademi lists; literary translation into and out of English, where a credited translator shares prize money and royalties.
Ed-tech, news apps and creator economy: regional-language scriptwriting for YouTube and short video, SEO content in Hindi/Tamil/Telugu, voice-over work, and vernacular news apps.

Interview questions

Q: What is the difference between translation, transcreation and localisation? A: Translation transfers meaning, transcreation rewrites the message for cultural effect in the target market, and localisation adapts the whole product — text, formats, numerals, images and layout — to the locale.
Q: How would you price a 40-minute video for subtitling into Malayalam? A: On video minutes, not words — rate per minute times 40, adjusted for whether a timed template is supplied, whether spotting is mine, and the turnaround; then check the effective hourly rate against my floor.
Q: Why is BLEU unreliable for Indian languages? A: It measures n-gram overlap with one reference, so free word order and rich morphology produce low scores even for correct output; chrF, COMET or human adequacy-fluency scoring are better.
Q: Calculate WER if a 200-word reference has 10 substitutions, 5 deletions and 5 insertions. A: (10+5+5)/200 = 20/200 = 10%.
Q: What does a kappa of 0.45 tell you about your annotation project? A: Only moderate chance-corrected agreement — the guidelines or label definitions are ambiguous, so revise them, retrain annotators and adjudicate before accepting the batch.
Q: Why should you never type Hindi in a Kruti Dev font for a client deliverable? A: It is a legacy font-encoded hack, not Unicode, so the text is unsearchable, breaks on other systems and cannot feed a translation memory or a model; use Unicode UTF-8.
Q: Who owns the copyright in a translation you produce for a publisher? A: Translation is a separate copyrightable work, but the publishing contract usually assigns or licenses it; the assignment must be in writing and the translator's moral rights, including credit, remain.
Q: A client gives you machine output and asks for "light post-editing" at 40% of your rate. How do you respond? A: Agree only after sampling the MT quality and measuring throughput; define light post-editing as fixing meaning, omissions and offensive errors only, and state that publication-quality output is full post-editing at a higher rate.
Q: What GST and TDS apply to a ₹50,000 translation invoice to an Indian company? A: 18% GST if I am registered, so ₹59,000 invoiced; the client deducts 10% TDS on the ₹50,000 fee (₹5,000), paying ₹54,000 and giving me a TDS credit.
Q: Which Act governs registration of a new Tamil monthly magazine? A: The Press and Registration of Periodicals Act, 2023, administered by the Press Registrar General of India; it replaced the Press and Registration of Books Act, 1867.
Q: How does NEP 2020 create jobs for a language graduate? A: It recommends mother-tongue/regional-language medium in early schooling and a three-language formula, and drives large-scale translation of higher-education material (the ASMITA project), expanding demand for teachers, translators and textbook editors.
Q: You must collect 100 hours of a dialect with no written standard. What is your first step? A: Write a transcription convention sheet with the orthography, code-mixing and disfluency rules, pilot it on a few hours with two annotators, measure agreement, then scale with gold items and QA sampling.

Common mistakes

Quoting a per-word rate without saying source or target words, or without the TM match bands — Indian-language targets expand and the invoice then disputes.
Comparing jobs by headline rate instead of effective hourly rate, so a high per-minute subtitling job with heavy spotting earns less than a plain translation.
Treating subtitles as a straight transcript: ignoring the character-per-line and reading-speed caps makes the file fail QC however good the language is.
Delivering in a legacy font encoding, or mixing Unicode and font-encoded text in one file.
Saying "machine translation has ended translation jobs" in an interview — the growth is in post-editing, LQA and data creation; the right answer names the shift, not the end.
Reporting raw percentage agreement as annotation quality instead of chance-corrected kappa, and skipping gold/honeypot items.
Forgetting that GST is charged on the fee while TDS is deducted on the fee only — and that export of services is zero-rated, not exempt, so an LUT and documentation are needed.
Signing a publishing contract without checking credit, royalty share and whether rights are assigned outright or licensed for a term and territory.