Indic Speech Data Collection: A Buyer’s Guide to Hindi, Hinglish and Indian-Language Voice Data

Last updat­ed: Octo­ber 2026

Voice prod­ucts for India fail in pre­dictable places: a caller switch­es from Hin­di to Eng­lish mid-sen­tence, an accent from east­ern Uttar Pradesh trips the rec­og­niz­er, or an agent built on clean stu­dio audio meets a noisy two-wheel­er street. Indic speech data col­lec­tion is how you close those gaps on pur­pose instead of by acci­dent.

This guide is for ASR, TTS and voice-agent teams buy­ing Hin­di, Hing­lish and Indi­an-lan­guage voice data. It cov­ers data types, dialect and demo­graph­ic sam­pling, code-switch­ing tran­scrip­tion, con­sent under Indi­a’s DPDP Act, cost dri­vers and a ven­dor score­card you can use in your next RFP.

Scop­ing an Indic voice dataset? Request a free 30-minute sam­ple set in your tar­get lan­guage or book a 20-minute scop­ing call.

What is Indic speech data collection?

Indic speech data col­lec­tion is the process of recruit­ing con­sent­ing speak­ers of Indi­an lan­guages, record­ing them to a defined spec­i­fi­ca­tion, and tran­scrib­ing and label­ing the audio so it can train or eval­u­ate speech mod­els. It cov­ers the 22 sched­uled lan­guages, region­al dialects, Hing­lish and oth­er code-switched speech, and Indi­an-accent­ed Eng­lish, deliv­ered with speak­er and record­ing meta­da­ta.

The work sits between lin­guis­tics and oper­a­tions. A dataset is only use­ful if the right peo­ple are speak­ing, the audio match­es your pro­duc­t’s chan­nel, the tran­scripts fol­low one con­ven­tion, and every file has a con­sent record behind it.

Three things make India hard­er than most mar­kets:

• Scale of vari­a­tion. The Con­sti­tu­tion’s Eighth Sched­ule lists 22 lan­guages, from Assamese to Dogri, per the Min­istry of Home Affairs. Each has region­al vari­eties that a sin­gle “lan­guage” label hides.
• Every­day code-switch­ing. Urban speak­ers mix Hin­di and Eng­lish (Hing­lish), or Tamil and Eng­lish, with­in one sen­tence. Mod­els trained on mono­lin­gual audio strug­gle with it.
• Script choic­es. The same spo­ken sen­tence can be writ­ten in Devana­gari, in Roman script or in a mix. Your tran­scripts must pick one con­ven­tion and hold to it.

In short: Indic speech data is con­sent­ed audio plus tran­scripts plus meta­da­ta, built to a spec that reflects how your users actu­al­ly speak. If you’re new to the idea of train­ing data in gen­er­al, our explain­er on what train­ing data is and how it shapes a mod­el cov­ers the basics.

Which types of Indic speech data do you need?

Most teams need a mix of five or six data types. Script­ed read speech gives clean, con­trolled cov­er­age for base­lines and TTS. Spon­ta­neous and two-speak­er con­ver­sa­tion­al speech reflects real usage for ASR and voice agents. Wake words, com­mands and nat­ur­al Hing­lish cov­er prod­uct-spe­cif­ic behav­ior. Indi­an-accent­ed Eng­lish sup­ports agents serv­ing Eng­lish-first callers.

Pick types based on what the mod­el will hear in pro­duc­tion, not on what’s eas­i­est to col­lect. Here’s how they map to use cas­es.

Data type	Main use cas­es	Col­lec­tion notes
Script­ed read speech	ASR base­lines, TTS voic­es, pro­nun­ci­a­tion cov­er­age	Pho­net­i­cal­ly bal­anced prompts in native script. Check read­ing flu­en­cy. TTS usu­al­ly needs stu­dio-grade audio from a small set of voic­es.
Spon­ta­neous (extem­pore) speech	ASR for real-world speech, speech mod­el pre­train­ing	Top­ic or pic­ture prompts pro­duce nat­ur­al paus­es, fillers and restarts. Tran­scrip­tion takes longer than for read speech.
Two-speak­er con­ver­sa­tion­al (call-cen­tre style)	Voice agents, con­tact-cen­tre ASR, diariza­tion, speak­er ID	Role-played sce­nar­ios such as bank­ing, tele­com or deliv­ery. Record each speak­er on a sep­a­rate chan­nel. Use fic­tion­al account details, nev­er real ones.
Wake words and com­mands	On-device key­word spot­ting, IVR, assis­tants	Many speak­ers say­ing a few phras­es, across dis­tances, devices and noise. Include near-miss neg­a­tives.
Code-switched Hing­lish	Voice agents and ASR for urban users	Elic­it nat­ur­al mix­ing through sce­nar­ios rather than forc­ing Eng­lish words into scripts. Agree the script and tag­ging con­ven­tion first.
Indi­an-accent­ed Eng­lish	Voice agents, glob­al ASR fine-tun­ing, accent eval­u­a­tion	Record each speak­er’s first lan­guage and home region as meta­da­ta, so accent groups can be bal­anced and test­ed.

Why voice agents need conversational, not just read, data

A voice agent has to han­dle turn-tak­ing, inter­rup­tions, hes­i­ta­tions, and callers who restate a request three ways. Read speech con­tains none of that. The IndicVoic­es project from AI4Bharat made the same choice for its open cor­pus: of its 7,348 hours, 9% was read, 74% extem­pore and 17% con­ver­sa­tion­al, accord­ing to the IndicVoic­es paper (ACL Find­ings 2024). For agent work, weight your bud­get toward the last two cat­e­gories, and ask for con­sent-backed con­ver­sa­tion­al speech data record­ed on the chan­nel your agent will use.

Indian accent English speech data

Many Indi­an callers start in Eng­lish, then switch. Indi­an accent Eng­lish speech data should be sam­pled by first lan­guage (a Tamil speak­er’s Eng­lish dif­fers from a Pun­jabi speak­er’s) and by region, not treat­ed as one accent. Include domain vocab­u­lary your agent will hear: prod­uct names, Indi­an place names, and amounts said in lakh and crore.

How do you plan language and dialect coverage?

Plan cov­er­age in lay­ers: pick lan­guages from your user base, then set quo­tas inside each lan­guage for dialect or region, age, gen­der, device and record­ing envi­ron­ment. Write the quo­tas as tar­gets with tol­er­ances, track them per batch, and cap min­utes per speak­er so a few voic­es don’t dom­i­nate the dataset.

Start with the offi­cial pic­ture. The Eighth Sched­ule’s 22 lan­guages are Assamese, Ben­gali, Gujarati, Hin­di, Kan­na­da, Kash­miri, Konkani, Malay­alam, Manipuri, Marathi, Nepali, Odia, Pun­jabi, San­skrit, Sind­hi, Tamil, Tel­ugu, Urdu, Bodo, San­thali, Maith­ili and Dogri. The list was last expand­ed in 2004, per the MHA doc­u­ment cit­ed above.

Then look under­neath the labels. The Cen­sus of India 2011 lan­guage report (Paper 1 of 2018) start­ed from 19,569 raw moth­er-tongue returns and ratio­nal­ized them into 121 lan­guages and 270 moth­er tongues with 10,000 or more speak­ers. The “Hin­di” group­ing alone, at 43.63% of the pop­u­la­tion, includes moth­er tongues such as Bho­jpuri, Rajasthani, Chhat­tis­garhi and Awad­hi. A Hin­di speech dataset record­ed only in Del­hi will under-serve a large share of peo­ple the cen­sus counts as Hin­di speak­ers.

Sampling quota planner

Dimen­sion	What to spec­i­fy	Exam­ple plan­ning choice
Dialect or region	States, dis­tricts or named vari­eties per lan­guage	Hin­di split across west­ern UP, Bihar (Bho­jpuri-influ­enced), Rajasthan and Mad­hya Pradesh
Age	Age bands and min­i­mum share per band	18–25, 26–40, 41–60, 60+. Minors only with ver­i­fi­able guardian con­sent.
Gen­der	Tar­get split with a tol­er­ance	Rough­ly bal­anced, with a tol­er­ance band agreed per batch
Device and chan­nel	Phone tiers, tele­pho­ny, head­set, far-field	Mix of bud­get and mid-range Android phones plus 8 kHz tele­pho­ny for agent work
Envi­ron­ment	Noise con­di­tions and their shares	Qui­et room, home with TV, street, vehi­cle, open office
Speak­er cap	Max­i­mum audio per speak­er	A per-speak­er minute cap so no voice dom­i­nates
Lit­er­a­cy	Whether read tasks are required	Read tasks need flu­ent read­ers. Spon­ta­neous tasks let you include speak­ers who don’t read the script.

Two prac­ti­cal rules. First, decide which dimen­sions are hard quo­tas (must hit) and which are soft (best effort). Hard-to-reach groups, such as speak­ers over 60 in rur­al dis­tricts, take longer and cost more, so mark them ear­ly. Sec­ond, keep a held-out test set with its own speak­ers. If the same voice appears in train­ing and test, your eval­u­a­tion num­bers will look bet­ter than real­i­ty.

How should Hinglish and code-switching speech data be transcribed?

Choose one script con­ven­tion (mixed script, full Devana­gari or full Roman), doc­u­ment it with exam­ples, and tag the lan­guage of each word or span. Add rules for loan­words, num­bers, named enti­ties and fillers. Apply one Uni­code nor­mal­iza­tion form. Then audit every batch against the guide­line, because code-switched tran­scripts drift fast.

Code-switch­ing speech data is where most Indic projects lose con­sis­ten­cy. Take one spo­ken sen­tence, “kal ki meet­ing resched­ule kar do” (“resched­ule tomor­row’s meet­ing”). It can be writ­ten three valid ways:

Con­ven­tion	Exam­ple	Works well for	Watch-outs
Mixed script	कल की meet­ing resched­ule कर दो	ASR that out­puts native script with word-lev­el lan­guage ID	Needs a rule for assim­i­lat­ed loan­words (ट्रेन or train?)
Full Devana­gari	कल की मीटिंग रीशेड्यूल कर दो	Hin­di-first ASR and TTS front ends	Eng­lish words get many Devana­gari spellings. A spelling lex­i­con is essen­tial.
Full Roman	kal ki meet­ing resched­ule kar do	Chat-style NLU that match­es how users type	No sin­gle stan­dard (“kar do” or “kar­do”). Spelling must be fixed by lex­i­con.

None of these is “cor­rect” in the abstract. The right choice is the one that match­es your mod­el’s out­put and down­stream use. What mat­ters is that every tran­scriber makes the same choice for the same word.

Tagging language spans

Script alone does­n’t tell you the lan­guage. “Phone” writ­ten in Devana­gari is still an Eng­lish-ori­gin word. So most code-switched specs add explic­it lan­guage tags, at one of two lev­els:

• Word lev­el: कल/hi की/hi meeting/en reschedule/en कर/hi दो/hi. This gives the most detail and sup­ports lan­guage-ID train­ing.
• Span lev­el: mark each con­tin­u­ous run of one lan­guage. It’s faster to anno­tate and enough for many ASR tasks.

Your guide­line should also set­tle these points before tran­scrip­tion starts:

• Loan­word list. Which Eng­lish-ori­gin words count as assim­i­lat­ed Hin­di (often “train”, “bus”, “phone”), and which get tagged as Eng­lish.
• Num­bers and amounts. Spo­ken form (“do baje”, “paanch lakh”) or dig­its. ASR train­ing usu­al­ly needs the spo­ken form.
• Named enti­ties. Brand, per­son and place names, with a fixed spelling list.
• Fillers and dis­flu­en­cies. Whether “mat­lab”, “haan” and false starts are kept, and how they’re marked.
• Nor­mal­iza­tion. One Uni­code nor­mal­iza­tion form for all files, and a con­sis­tent nuk­ta pol­i­cy (for exam­ple, ज़ ver­sus ज), so iden­ti­cal words don’t become dif­fer­ent tokens.
• Roman­iza­tion scheme. If you use Roman script, say whether it is infor­mal “chat” spelling or a for­mal scheme such as ISO 15919.

The same log­ic applies to Tang­lish (Tamil-Eng­lish), Beng­lish and oth­er pairs, and to Urdu, which uses Per­so-Ara­bic script while shar­ing much of its spo­ken base with Hin­di. For more on how tran­scripts feed recog­ni­tion mod­els, see our guide to AI tran­scrip­tion, accu­ra­cy and WER.

What consent and compliance does Indic voice data need?

Every speak­er should give explic­it, informed con­sent in a lan­guage they under­stand, for named pur­pos­es such as ASR train­ing or syn­thet­ic voice cre­ation, with an easy way to with­draw. In India this aligns with the DPDP Act 2023. For speak­ers in the EU, GDPR applies, and voice used to iden­ti­fy peo­ple can be spe­cial-cat­e­go­ry bio­met­ric data.

Con­sent­ed voice data for AI is more than a signed form. It’s a chain of evi­dence link­ing each audio file to a spe­cif­ic per­son­’s informed agree­ment for a spe­cif­ic use.

India: DPDP Act 2023 and DPDP Rules 2025

The Dig­i­tal Per­son­al Data Pro­tec­tion Act, 2023 (No. 22 of 2023) requires that con­sent be “free, spe­cif­ic, informed, uncon­di­tion­al and unam­bigu­ous with a clear affir­ma­tive action” (Sec­tion 6). It also requires that with­draw­ing con­sent be as easy as giv­ing it.

The DPDP Rules, 2025 were pub­lished as G.S.R. 846(E), dat­ed 13 Novem­ber 2025, and announced by MeitY the next day. They com­mence in phas­es. As of Octo­ber 2026, per the Rules’ com­mence­ment clause:

• Rules 1, 2 and 17 to 21 (def­i­n­i­tions and the Data Pro­tec­tion Board) took effect on pub­li­ca­tion.
• Rule 4 (con­sent man­agers) takes effect one year after pub­li­ca­tion, in Novem­ber 2026.
• Rules 3, 5 to 16, 22 and 23 (includ­ing notice, secu­ri­ty safe­guards, breach report­ing and chil­dren’s data) take effect eigh­teen months after pub­li­ca­tion, around May 2027.

Indus­try reports in 2026 said MeitY had con­sult­ed on short­en­ing parts of this time­line. Check the cur­rent noti­fi­ca­tion before you final­ize your com­pli­ance plan. Either way, build to the 2027 stan­dard now. Datasets you col­lect today will still be in use when the full rules apply.

EU speakers: GDPR

If you record speak­ers locat­ed in the EU, the GDPR is like­ly to apply. Treat voice record­ings as per­son­al data. Where voice is processed to unique­ly iden­ti­fy some­one (speak­er ID or voice bio­met­rics), it becomes bio­met­ric data under Arti­cle 9’s spe­cial cat­e­gories. Pro­cess­ing it is then pro­hib­it­ed unless an excep­tion applies, and for data col­lec­tion that excep­tion is usu­al­ly explic­it con­sent. Take legal advice on your spe­cif­ic set­up.

What a consent record should hold

• Pseu­do­ny­mous par­tic­i­pant ID linked to every file they con­tributed
• The notice shown, its ver­sion and its lan­guage
• Each pur­pose con­sent­ed to, as sep­a­rate items (ASR train­ing, eval­u­a­tion, TTS, voice cloning)
• Who will receive the data, and how long it will be retained
• How the par­tic­i­pant can with­draw, and what hap­pens to their files if they do
• Time­stamp and form of the affir­ma­tive action
• Guardian details and ver­i­fi­ca­tion for any minor

Pur­pose lim­i­ta­tion in prac­tice: con­sent to “speech recog­ni­tion research” does­n’t cov­er cloning some­one’s voice. If TTS or syn­thet­ic voice is in scope, ask for it as a sep­a­rate, clear­ly explained opt-in. Also keep per­son­al data out of the audio. Use fic­tion­al names and account num­bers in role-plays, and redact any per­son­al details that come up in spon­ta­neous speech.

How does an Indic speech data collection project run?

A well-run project moves through six stages: a writ­ten spec­i­fi­ca­tion, speak­er recruit­ment with con­sent, man­aged record­ing, tran­scrip­tion and anno­ta­tion, qual­i­ty assur­ance, and deliv­ery with full meta­da­ta. A short pilot through all six stages, in one lan­guage, catch­es most prob­lems before you com­mit to full vol­ume.

1. Spec­i­fi­ca­tion. Lan­guages, hours, data types, quo­tas, audio for­mat (for exam­ple, 16 kHz 16-bit mono WAV for ASR, 8 kHz for tele­pho­ny, 48 kHz for TTS), the tran­scrip­tion con­ven­tion and the accep­tance cri­te­ria.
2. Speak­er recruit­ment and con­sent. Native-speak­er recruiters source par­tic­i­pants against the quo­tas. Each per­son gets the notice in their lan­guage, gives con­sent per pur­pose, and is screened for lan­guage flu­en­cy and dupli­cate enroll­ment.
3. Record­ing man­age­ment. Prompts and sce­nar­ios are served through a record­ing app or call set­up. Auto­mat­ic checks flag clip­ping, silence, wrong sam­ple rate and back­ground noise at cap­ture time, so speak­ers can re-record on the spot.
4. Tran­scrip­tion and anno­ta­tion. Native-speak­er tran­scribers fol­low the writ­ten guide­line: script con­ven­tion, lan­guage tags, time­stamps, speak­er turns, and noise and event tags. Our audio tran­scrip­tion ser­vices page cov­ers the ver­ba­tim and time-aligned options.
5. Qual­i­ty assur­ance. A sec­ond review­er audits a sam­ple (or every file, for eval­u­a­tion sets), with errors logged by type. Batch­es below the agreed thresh­old go back for rework.
6. Deliv­ery with meta­da­ta. Audio, tran­scripts and a man­i­fest with speak­er ID, lan­guage, dialect or region, age band, gen­der, device, envi­ron­ment and con­sent ref­er­ence for each file.

How much does speech data collection cost per hour?

Speech data col­lec­tion cost per hour depends on lan­guage rar­i­ty, how spon­ta­neous the speech is, unique speak­ers per hour, audio spec, record­ing envi­ron­ment, tran­scrip­tion depth, QA lev­el and licens­ing terms. Read speech in a wide­ly spo­ken lan­guage with light tran­scrip­tion sits at the low end. Exclu­sive con­ver­sa­tion­al data in a rare dialect sits at the high end.

Instead of trust­ing a sin­gle per-hour fig­ure, ask any ven­dor to price against your spec and to show how each dri­ver below moves the num­ber.

Cost dri­ver	Push­es cost down	Push­es cost up
Lan­guage rar­i­ty	Hin­di, Ben­gali, Marathi, Tamil, Tel­ugu	Small­er lan­guages and named dialects with few avail­able speak­ers
Speech type	Script­ed read speech	Spon­ta­neous and two-speak­er con­ver­sa­tion, which take more set­up and hard­er tran­scrip­tion
Speak­ers per hour	Few­er speak­ers record­ing longer ses­sions	Many unique speak­ers, each con­tribut­ing a few min­utes
Audio spec­i­fi­ca­tion	16 kHz mono on smart­phones	48 kHz stu­dio cap­ture, mul­ti-chan­nel or far-field arrays
Envi­ron­ment	Speak­ers’ own homes, uncon­trolled	Stu­dio booths or spe­cif­ic con­trolled noise con­di­tions
Tran­scrip­tion depth	Ortho­graph­ic tran­script only	Ver­ba­tim with time­stamps, lan­guage-span tags, diariza­tion and event labels
QA lev­el	Sam­ple-based audit	100% sec­ond-pass review and dou­ble tran­scrip­tion for test sets
Demo­graph­ic quo­tas	Loose tar­gets	Hard quo­tas for hard-to-reach groups, such as rur­al speak­ers over 60
Licens­ing	Non-exclu­sive licence	Exclu­sive rights or full own­er­ship trans­fer

Two levers usu­al­ly save the most mon­ey with­out hurt­ing qual­i­ty. First, tran­scribe only what you’ll train on. Spon­ta­neous audio can be col­lect­ed at vol­ume and tran­scribed in phas­es. Sec­ond, sep­a­rate the eval­u­a­tion set. Spend on 100% QA for a small­er, care­ful­ly bal­anced test set, and use sam­ple-based QA for bulk train­ing data.

How to choose an Indic speech data vendor: an 8‑point scorecard

Score each ven­dor from 0 to 2 on eight checks: a real sam­ple in your lan­guage, con­sent arte­facts, speak­er prove­nance, quo­ta report­ing, a writ­ten tran­scrip­tion guide­line, mea­sur­able QA, audio spec checks and clear licens­ing. A score of 14 to 16 means ready to pilot. 10 to 13 means pilot with con­di­tions. Below 10, keep look­ing.

1. Sam­ple in your lan­guage. A real sam­ple set in your tar­get lan­guage and data type, with tran­scripts and a meta­da­ta man­i­fest. Not a gener­ic demo reel.
2. Con­sent arte­facts. The par­tic­i­pant notice in the local lan­guage, the per-pur­pose con­sent screen, the with­draw­al process, and con­sent records linked to file IDs.
3. Speak­er prove­nance. Real, unique speak­ers with dupli­cate checks. Any syn­thet­ic or aug­ment­ed audio is dis­closed and labeled.
4. Quo­ta report­ing. Per-batch reports of actu­al dialect, age, gen­der, device and envi­ron­ment shares against your tar­gets.
5. Writ­ten tran­scrip­tion guide­line. Cov­ers script con­ven­tion, lan­guage tags, loan­words, num­bers, fillers and nor­mal­iza­tion, with worked exam­ples.
6. Mea­sur­able QA. A defined met­ric (such as tran­script error rate on an audit­ed sam­ple), the sam­ple size, rejec­tion rea­sons and rework rules.
7. Audio spec checks. Auto­mat­ed checks for sam­ple rate, bit depth, clip­ping, silence, sig­nal-to-noise and chan­nel sep­a­ra­tion.
8. Licens­ing and data han­dling. Clear rights grant, exclu­siv­i­ty terms, stor­age loca­tion, reten­tion and dele­tion, and any sub-proces­sors.

Add one com­mer­cial test on top: does the ven­dor invoice only for files that pass your accep­tance cri­te­ria? Pay­ing for accept­ed files keeps both sides focused on the same qual­i­ty bar.

Want to run this score­card on us? Request a free 30-minute sam­ple set in your tar­get lan­guage and score it against your own spec, or scope a pilot billed only on approved files.

Common mistakes in Indic speech data projects

• Treat­ing “Hin­di” as one accent. With­out dialect and region quo­tas, the dataset skews to who­ev­er was eas­i­est to recruit.
• Forc­ing code-switch­ing into scripts. Script­ed Hing­lish sounds unnat­ur­al. Use sce­nar­ios that make peo­ple mix lan­guages the way they nor­mal­ly do.
• Stu­dio audio for a phone prod­uct. If users call over tele­pho­ny, record and test at tele­pho­ny qual­i­ty.
• Chang­ing the tran­scrip­tion con­ven­tion mid-project. It splits your data into incon­sis­tent halves. Lock it after the pilot.
• Vague con­sent. A sin­gle “AI research” check­box won’t cov­er voice cloning or resale. Name each pur­pose.
• Speak­er over­lap between train­ing and test sets. It inflates eval­u­a­tion results and hides real-world errors.

How Graveiens AI delivers Indic speech data

Graveiens AI runs Indic voice projects on an explic­it-con­sent work­flow: par­tic­i­pant onboard­ing → record­ing man­age­ment → meta­da­ta tag­ging → QA. Every con­trib­u­tor is onboard­ed with con­sent before record­ing, every file car­ries meta­da­ta, and every batch pass­es QA before deliv­ery. You’re invoiced only for approved files, which keeps a first pilot low-risk.

Indic lan­guages we cov­er: Hin­di, Urdu, Ben­gali, Marathi, Gujarati, Pun­jabi, Assamese, Tel­ugu, Tamil, Malay­alam, Kan­na­da, Odia, Konkani and Nepali, plus Indi­an Eng­lish. These sit with­in our wider cov­er­age of 25+ lan­guages, and both script­ed and con­ver­sa­tion­al col­lec­tion are avail­able on our voice and speech data ser­vices page.

How a typ­i­cal engage­ment runs:

• Scop­ing. We turn your use case into a writ­ten spec cov­er­ing data types, quo­tas, audio for­mat, the code-switch­ing con­ven­tion and accep­tance cri­te­ria. For any­thing out­side our stan­dard cat­a­logue, we scope it with you.
• Pilot. One lan­guage and one data type, end to end, so you can check audio, tran­scripts and meta­da­ta against your spec before scal­ing.
• Scale-up. Batch-wise deliv­ery with quo­ta report­ing, tran­scrip­tion by native speak­ers, and the same QA gate on every batch.
• Lan­guage sup­port. Trans­la­tion, local­iza­tion of prompts and scripts, and sub­ti­tling are avail­able through our lan­guage and local­iza­tion ser­vices.

Graveiens AI is ISO 9001 cer­ti­fied. We don’t promise accu­ra­cy gains for your mod­el. That depends on your archi­tec­ture and eval­u­a­tion. What we com­mit to is data that match­es the spec you sign off.

Sources

1. Min­istry of Home Affairs, Gov­ern­ment of India. Lan­guages includ­ed in the Eighth Sched­ule of the Con­sti­tu­tion.
2. Office of the Reg­is­trar Gen­er­al and Cen­sus Com­mis­sion­er, India. Lan­guage (Paper 1 of 2018), Cen­sus of India 2011.
3. Min­istry of Law and Jus­tice. The Dig­i­tal Per­son­al Data Pro­tec­tion Act, 2023 (No. 22 of 2023), offi­cial text host­ed by MeitY.
4. Javed et al. IndicVoic­es: Towards build­ing an Inclu­sive Mul­ti­lin­gual Speech Dataset for Indi­an Lan­guages. Find­ings of ACL 2024.
5. Min­istry of Elec­tron­ics and Infor­ma­tion Tech­nol­o­gy. Dig­i­tal Per­son­al Data Pro­tec­tion Rules, 2025, G.S.R. 846(E), Gazette of India, 13 Novem­ber 2025 (pub­lished on meity.gov.in).
6. Dig­i­tal India Cor­po­ra­tion, MeitY. Bhashi­ni, Nation­al Lan­guage Trans­la­tion Mis­sion (bhashini.gov.in).