Multilingual Forms: Let People Answer in Hindi, Tamil or Any Language They Think In

A form written in English reaches fewer Indians than you think, and a form written in Hindi does not fix it. Hindi is the mother tongue of 43.63% of the country, according to the 2011 Census. A multilingual survey has to deal with a simple fact: people often read in one language, think in a second, and type badly in both.

People think in one language and type in another. That gap, not the translation of your questions, is where most multilingual surveys lose their best answers. This post sets out what the census says about how many languages your respondents carry, why typing them is hard, and the three ways to let people answer in their own language.

Most of your respondents carry more than one language

The 2011 Census is still the latest published language data for India. It asked every person for their mother tongue, and the answers were messy in an instructive way. Enumerators recorded 19,569 raw mother tongue names, which the Registrar General grouped into 121 languagesCensus of India 2011, Language, Paper 1 of 2018, General Note, paragraphs 4 and 5. Office of the Registrar General, India.. Of those, 22 are the Scheduled Languages listed in the Constitution.

The 22 Scheduled Languages cover 96.71% of the population as a mother tongue. That sounds tidy until you look at the spread. No single language is a majority, and the fifth largest, Tamil, is still the first language of 69 million people.

Hindi is a bigger tent than it looks

Even the 43.63% overstates how many people speak standard Hindi at home. The census groups dozens of mother tongues under Hindi. Of the 528 million people counted under Hindi, about 322 million returned "Hindi" itself. Another 50.6 million returned Bhojpuri and 25.8 million returned Rajasthani.

That matters for a survey. Someone whose home language is Bhojpuri can read a Hindi form. They may still explain a problem with their crop or their landlord more freely in the language they grew up with. A form that only accepts formal Hindi text quietly asks them to translate themselves.

One in four Indians reported a second language

The census also records subsidiary languages. We summed the national C-17 table. About 315 million people, 26% of the population, reported at least one second language, and about 86 million, 7%, reported a thirdCensus of India 2011, Table C-17, Population by bilingualism and trilingualism, India. Our sum across all mother tongues: 314,988,770 second-language speakers and 86,009,580 third-language speakers, of 1,210,854,977.. Those are self-reports, so they measure what people claim, not tested fluency.

English is the case that should worry anyone building forms for India. In the same table, only about 260,000 people gave English as their mother tongue. Roughly 83 million reported it as a second language and 46 million as a third. Taken together, that is about 1 in 10 Indians who report English at all.

The bilingual share depends heavily on the language

Bilingualism is not spread evenly. Our calculation from the C-17 table shows large gaps between mother tongues. Of the large groups we checked, Hindi speakers were the least likely to report a second language, which makes sense when your language is the one others learn.

We read this chart as a warning against a common shortcut. Teams often assume that a Hindi form plus an English form covers India. For a Tamil respondent, both may be second or third languages, and for many, neither appears at all.

Typing an Indian language is its own barrier

Even when your respondent reads the form's language well, typing an answer in it is a separate skill. This is the part most multilingual survey advice skips. Reading Tamil script is easy for a Tamil speaker. Typing a long answer in it on a phone keyboard is often not.

Most people type Indian languages in the Latin alphabet, with no fixed spelling

Researchers at Google spelled out the problem when they released the Dakshina dataset. South Asian languages are often written in the Latin script for informal communication such as SMS, WhatsApp and social media, and they lack a common standard romanisationRoark et al., Processing South Asian Languages Written in the Latin Script: the Dakshina Dataset, LREC 2020, Introduction. Dataset covers 12 South Asian languages.. Their example is the Hindi word for corruption. It turned up romanised as bhrastachar, bhrashtachar, barashtachaar and bharastachar, in a single comment thread.

A separate Google Research paper on phone keyboards says the same in its abstract: Latin-script text entry for South Asian languages is common, with no standard spelling (Wolf-Sonkin and colleagues, 2019). Neither paper gives a share of users who type this way. We could not find a primary source that does, so we will not quote one.

What this does to your survey data

Romanised answers are readable, but they are hard to search, count or translate. "Paani nahi aata" and "pani nahin ata" are the same complaint, and a keyword filter treats them as two. The Dakshina authors also note that this script variance is a challenge for machine translation and information retrieval, which is what you would use to analyse the answers.

Native-script typing solves the data problem and creates a respondent problem. People need the right keyboard installed and the patience to use it. On a long open question, many will shorten their answer to what is easy to type. We wrote about how typing shapes open-ended answers and the effect is stronger in a second script.

Three ways to let people answer in their own language

There are only three real options. Most good multilingual forms combine two of them.

The last row is a concession worth taking seriously. Voice is the wrong tool for a phone number or a district name. Keep those as typed or tapped fields in every language.

Option 1: translate the questions properly

Translation fixes comprehension, and it is still the right first step for the question text. The trap is treating it as a one-pass job. The European Social Survey uses a team method called TRAPD: translation, review, adjudication, pretesting and documentation. It aims for equivalent meaning, not word-for-word copies.

You do not need a committee for a customer survey. You do need one native speaker to read every translated question, and a small pilot in each language. Rating scales are the usual casualty. A literal translation of "somewhat satisfied" can land as far more formal, or far more negative, than you meant.

Option 2: accept answers in any script

This option costs nothing. Most text fields in any modern form builder accept Unicode, so a respondent can type in Devanagari, Tamil script, Bengali or romanised text. The form does not care.

The cost lands later. You get a mix of scripts and spellings in one column, and short answers from people who found typing tiring. We think this is acceptable for short answers and a poor default for anything where you want a story.

Option 3: let them speak

Speaking removes both barriers at once. The respondent hears or reads the question, taps a microphone and answers in whatever language comes naturally. That includes the Hindi and English mix many urban respondents actually use. We covered why code-switching shows up in voice answers in more depth.

Voice has a real cost: someone has to listen to the answers, or transcribe and translate them. It also loses people who are somewhere they cannot talk. Our analysis of spoken survey answers covers that drop-off and how to design around it. We still think voice is the best default for open questions in a multilingual survey, because it moves the language burden from the respondent to you.

What Would You Rather does today, and what it does not

We build Would You Rather, so here is a plain account of what it handles for multilingual forms. Everything below was checked against the product's code in October 2026.

What works for multilingual forms

You can write questions and answer options in any script, because they are plain text. You can attach an audio clip to any question, so the question plays aloud. That helps when a respondent reads the language poorly, or when you want the question in a regional dialect with no settled spelling.

The voice response field records in the phone browser, with no app. It accepts any spoken language, because it stores audio, not text. You can set a minimum and maximum length per question. An optional "or type your answer" box can sit under the microphone. It is off by default, and you switch it on per question.

Transliteration suggestions, in one specific mode

Would You Rather has an on-screen transliteration helper. The respondent types a word in English letters, for example "namaskaram", and a strip of native-script suggestions appears below the box. Tapping one replaces the word. It works on any phone, without an Indian-language keyboard installed. The suggestions come from Google's public transliteration service, and the language is set per question.

Be clear about its scope. Today it appears only in a transcript mode used on our speech-data projects, where a respondent records a sentence and corrects its written version. It is not yet a switch in the form editor for ordinary text questions. If you need it on a form, it is set up for you, not toggled by you. We describe that workflow in building a speech dataset for ASR.

What it does not do

Would You Rather does not translate your form. It does not translate answers. It does not transcribe voice recordings into text. The respondent interface, meaning buttons like Submit and the small prompts around fields, is in English. Your question text can be in any language, but the frame around it is not.

If you need English text from Tamil recordings, you export the audio links and run them through a speech recognition or translation tool. Then you check a sample with someone who speaks Tamil. We would not quote any machine transcript of an Indian language without that check.

How to set up a multilingual form, step by step

This is the order we recommend. It works in any form builder with a voice field, and the language choices apply even without one.

If your audience includes people who read slowly or not at all, the design changes further. We cover that in survey tools for low-literacy respondents. For forms filled in villages and markets, our notes on field research forms cover offline gaps, shared phones and QR codes.

Where this argument is weakest

The census figures are 15 years old. The 2011 tables are the latest published language data, but migration, schooling and smartphones have all shifted since. English and Hindi as second languages have probably grown, especially among younger urban respondents, and we cannot say by how much.

Bilingualism figures are self-reported. A person who says they speak English may read it well and type it poorly, or the reverse. The census does not test skill, and it records at most two subsidiary languages per person.

The typing evidence is qualitative. Google's researchers describe romanised typing as common, but neither paper we opened gives a share of users. We have no first-party data comparing typed and spoken answers by language on our own forms, so we make no claim about how much longer Tamil voice answers run than Tamil typed ones.

The strongest counter-argument is about cost. Voice answers in five languages need five sets of ears, or a translation pipeline you trust. For a team without those, a carefully translated form with choice questions may produce more usable data than a pile of recordings no one can understand. That is a fair choice. It just should not be the default for questions where you want people to explain.

Where to start

Pick one open question from your current survey, the one where answers are shortest. Translate it into the second most common language in your audience, add a voice option, and say "answer in any language" in the question. Send it to 20 people in that group.

Then listen to the answers before you decide anything. If you hear details you never got in typed replies, expand language by language. If you hear silence, keep the text box and fix the question. Our guide to adding audio questions to forms covers the setup.

References

1. Office of the Registrar General, India, Census of India 2011, Language, India, States and Union Territories (Table C-16), Paper 1 of 2018. Used for mother tongue shares, the 19,569 raw returns, the 121 languages, the 96.71% Scheduled Language coverage and the mother tongues grouped under Hindi.
2. Office of the Registrar General, India, Census of India 2011, Table C-17: Population by bilingualism and trilingualism, India. Used for second and third language counts, including English. The totals and per-language shares are our sums of the published table.
3. Roark, B., Wolf-Sonkin, L., Kirov, C., Mielke, S. J., Johny, C., Demirşahin, I. and Hall, K., Processing South Asian Languages Written in the Latin Script: the Dakshina Dataset, LREC 2020. Used for informal Latin-script writing and the lack of standard romanisation.
4. Wolf-Sonkin, L., Schogol, V., Roark, B. and Riley, M., Latin script keyboards for South Asian languages with finite-state normalization, FSMNLP 2019. Used for the prevalence of Latin-script text entry.
5. European Social Survey, Translation, ESS methodology pages, accessed 8 October 2026. Used for the TRAPD translation method.
6. Would You Rather source code, checked 8 October 2026. Used for the voice field, text fallback, audio question prompts and the scope of transliteration suggestions.

The weakest point in this source base: the census data is from 2011 and self-reported, and no primary source we opened puts a number on how many Indians type their language in romanised text.