Code-Switching in Indian Voice Data: What 2,731 Scripts and 805 Typed Answers Show

We set out to measure code-switching in voice answers and hit a wall first. Would You Rather does not transcribe recordings, so we cannot say how often respondents mix Hindi and English when they speak. What we can measure turned out to be more revealing. Of 2,185 Indian-language sentences that speech-data teams asked people to read aloud on our platform, 90% deliberately mixed in English words.

The people buying Indian speech data are asking for code-switched speech, on purpose, at scale. That tells you what real-world voice systems need to understand. We also found that when people type an Indian language, nearly half reach for Latin letters. Both findings come from read-only queries on our production database, with published research on code-switched speech recognition for context.

What we could measure, and what we couldn't

Every first-party number here comes from Would You Rather's production database, queried read-only on 8 October 2026, covering 18 April to that date. We publish aggregates only: no sentences, no answers and no form names.

The honest gap comes first. Code-switching in the audio itself would need transcripts and language identification for every recording. We have neither. Would You Rather stores and plays recordings but does not transcribe them, and we did not run a speech recogniser for this post. So we measured the text around the voice instead.

Two kinds of text we could analyse

The first is the prompts. On script-reading forms, a speech-data team writes sentences and respondents read them aloud. Those sentences show what language mix the buyers of speech data want recorded. The second is typed answers. On one large application form, respondents typed a transcription of a short clip in the language they chose.

For both, we classified every word by the writing system of its letters: Latin, or one of the Indian scripts such as Devanagari, Tamil or Malayalam. A text is "mixed" if it contains words from both. That is a script measure, not a language measure, and the difference matters below.

Which languages people recorded in

Before the mixing, the volume. Script-reading forms on our platform produced 4,467 recordings. By the language each question was tagged with, or the script its sentence was written in, about 1,711 were Hindi sentences in Devanagari and 1,032 were Malayalam. Tamil added 411, Odia 120 and Assamese 113, with a few dozen in Telugu and Kannada. 473 were tagged English, and 573 came from untagged sentences in Latin letters.

The application form shows the spoken side. It asked people to pick a first-preference language and then explain an everyday process aloud in that language. Of 918 completed applications, 457 chose Hindi, 198 Tamil, 95 Telugu, 69 Malayalam and 61 Kannada. Marathi, Bengali, Odia and Punjabi together accounted for 29, and 9 left the question blank.

896 own-language recordings we can't yet read

That form holds 896 recordings of people explaining a process in their own language, unscripted. It is exactly the material in which code-switching would show up: everyday instructions, in a regional language, from people who also use English at work. Without transcripts, we can't count a single switch in it.

We think this is the most useful next measurement for anyone building Indian voice products, and we would rather name the gap than fill it with a guess. A sample of 100 to 200 of those recordings, transcribed by native speakers with each word tagged by language, would answer the question this post's title asks. Until then, the prompt and typing data are the best evidence we have.

90% of Indian-language scripts mixed in English

Across our script-reading forms we found 2,731 distinct sentences. 546 were entirely in Latin letters, mostly English prompts. Of the 2,185 written in an Indian script, 1,961, or 90%, contained at least one English word in Latin lettersWould You Rather, 31 script-reading forms, distinct prompt sentences deduplicated by text. 1,961 of 2,185 Indian-script prompts contained at least one Latin-script word. Median Latin-word share in mixed prompts 18.5%..

The English was a seasoning, not the main dish. In the median mixed sentence, about 18.5% of words were in Latin letters. The pattern is familiar to anyone who has spoken to a customer service line in India: an Indian-language sentence carrying English nouns like order, refund, branch or process.

Smaller sets told the same story. All 10 Odia sentences and all 10 Assamese sentences mixed in English words. Those are too few to call a rate, but nobody wrote a pure-script sentence in either language.

Our count understates the mixing

Script is a blunt proxy for language. An English word written in Devanagari letters counts as Hindi in our method, and we saw that happen. A Hindi word typed in Latin letters would count as English. We saw examples of the first case in the Hindi prompts, so treat 78% as a floor, not an exact rate.

Why buyers write mixed scripts

We think the reason is simple. Speech-data buyers want models that understand how people actually talk, and people in India talk in a mix. A recogniser trained only on formal, single-language Hindi will stumble on a sentence like "aapka refund process ho gaya hai". The scripts on our platform read like real support calls, and that is deliberate.

That is consistent with older research on written Hindi online. In a 2014 study of English-Hindi Facebook posts by Bali and colleagues at Microsoft Research India, 23.7% of Hindi sentences contained at least one English word, against 7.2% of English sentences with a Hindi wordBali, Sharma, Choudhury and Vyas, First Workshop on Computational Approaches to Code Switching, ACL 2014. 381 Facebook posts after filtering, 4,135 words, annotated by linguists.. That sample was small and from one social platform, but the direction matches: English goes into Hindi far more than the other way round.

Nearly half typed Hindi in Latin letters

The second dataset is typed text. On one application form, 918 people completed a section that played a short clip in their chosen language and asked them to transcribe it. 805 typed something. We classified each answer by script.

Overall, 328 answers were typed entirely in an Indian script, 287 entirely in Latin letters and 186 in a mix. Four contained no letters we could classify. For Hindi, the largest group, 199 of 421 people typed the Hindi clip entirely in Latin letters, more than the 183 who used DevanagariWould You Rather, one application form, 421 non-empty Hindi transcription answers. Latin only 199; Devanagari only 183; mixed 36; no classifiable letters 3. Queried 8 October 2026..

The pattern differed sharply by language

Hindi typists split almost evenly between Devanagari and Latin letters. Telugu and Malayalam typists mostly used their own scripts. Tamil answers were mostly mixed. We do not know what each clip contained, so a clip with English words in it would push its language toward "mixed" for faithful reasons.

Romanised typing has no standard spelling

This matters for anyone storing typed answers alongside recordings. Google researchers describe the same habit in their Dakshina dataset paper. South Asian languages are often written in Latin script for informal communication, native-script keyboards are less common than Latin input, and there is no standard romanisation. The same Hindi word can be spelled several ways by different people.

Why code-switched speech is still hard to transcribe

If your voice answers are mixed, your transcription engine has to cope. The best public benchmark we found for Indian code-switched speech is the MUCS 2021 challenge, published at Interspeech by Diwan and colleagues. It released about 600 hours of transcribed speech across seven Indian languages, including Hindi-English and Bengali-English code-switched sets.

The baseline systems for the code-switched task, averaged across the two pairs, scored word error rates between about 28% and 34% depending on the model and test set. Roughly one word in three wrong is not good enough to quote from without checking. Those were 2021 baselines, and newer systems do better, but the challenge it measured has not gone away: two languages, two scripts and very little mixed training data.

What that means for a voice form

We would not trust any automatic transcript of a Hinglish voice answer without spot-checking it by ear. Decide which script your transcripts should use for English words before anyone transcribes anything. Mixed conventions inside one dataset are worse than either convention used consistently.

The same applies to typed answers you plan to analyse. If nearly half of your Hindi respondents type in Latin letters with no standard spelling, a keyword search for a Devanagari word will miss them, and a search for one romanised spelling will miss the others. Before you count themes in open answers from Indian respondents, normalise them into one script, or have a reader who handles both code them by hand. A dashboard that only reads Devanagari will quietly under-count the people who typed the way they text.

Where this analysis is weakest

The prompt data reflects what a few speech-data projects wrote, not how people talk unprompted. Almost all of it comes from forms owned by one account, and a handful of projects contributed most of the sentences. A different buyer might write purer scripts.

The typed data comes from one form, from job applicants who chose their language from a list. Script choice when transcribing a clip may differ from script choice when writing freely. And because we never saw the clips, we cannot separate people who mixed scripts by habit from people who faithfully copied English words they heard.

Most importantly, none of this measures code-switching in speech. It measures scripts in text near the speech. A proper study would transcribe a sample of voice answers, label each word's language, and report how often speakers switch. We have not done that yet.

How to plan a voice form for mixed-language speakers

If your respondents speak Indian languages, assume mixing and design for it. The steps below are the ones we would follow on our own forms.

On speech-data forms, Would You Rather has a transcript mode that shows the reading sentence under the recorder in an editable box. Respondents can type in English letters and get suggestions in their own script, which helps people who don't have an Indian-language keyboard. For more on the collection side, read building a speech dataset for ASR and the speech data collection tools we compared. For surveys rather than datasets, see collecting answers in Hindi, Tamil or any language and what 6,274 voice answers on our platform actually were.

References

1. Diwan, A. and colleagues, MUCS 2021: Multilingual and Code-Switching ASR Challenges for Low Resource Indian Languages, Interspeech 2021. Used for code-switched ASR data and baseline word error rates.
2. Roark, B. and colleagues, Processing South Asian Languages Written in the Latin Script: the Dakshina Dataset, LREC 2020. Used for romanised writing of South Asian languages and the lack of a standard romanisation.
3. Bali, K., Sharma, J., Choudhury, M. and Vyas, Y., "I am borrowing ya mixing?" An Analysis of English-Hindi Code Mixing in Facebook, First Workshop on Computational Approaches to Code Switching, ACL 2014. Used for English-in-Hindi mixing rates in written posts.
4. Would You Rather, production database, read-only queries, 8 October 2026. Used for every first-party figure. Aggregated and anonymised; not independently audited.

The weakest point in this source base: our first-party data measures scripts in text, not languages in speech, and the published ASR figures are 2021 baselines rather than the best systems available today.