Hinglish Voice AI: How Code-Switching Works & How to Test It

Code-switching is the practice of alternating between two or more languages within a single conversation, or even a single sentence. In Hinglish, a caller might say, 'Mera order abhi tak deliver nahi hua, can you check the status?' A voice agent handles this well only if its speech recognition, language understanding, and speech synthesis all treat mixed-language speech as normal input rather than as an error.

For Indian businesses this is not an edge case. It is how many callers actually talk on the phone, and you can hear how much of it your own callers do in your call recordings. This guide explains the mechanics, lists what a Hinglish-capable agent has to get right, and gives you a 10-point protocol you can run on any vendor, including Vaami, using your own audio.

What is code-switching, linguistically?

Linguists distinguish two common patterns. Inter-sentential switching moves between languages at sentence boundaries: a full sentence in Hindi, then a full sentence in English. Intra-sentential switching, the harder case for machines, mixes languages inside one sentence, usually keeping one language's grammar as the frame and dropping in words from the other.

In everyday Hinglish the grammatical frame is usually Hindi and the borrowed content is English: nouns like 'appointment', 'balance', or 'delivery', plus fixed phrases like 'no problem' or 'please confirm'. The English words then take Hindi grammar around them: 'appointment cancel kar do', 'EMI ka date change ho sakta hai?' The English noun stays English, but the verb, postposition, and word order follow Hindi.

• Matrix language: the language that supplies the sentence's grammatical frame (often Hindi in Hinglish).
• Embedded language: the language supplying inserted words or phrases (often English).
• Speaker variation: the Hindi-to-English ratio changes by person, city, topic, and even mood. A caller may be nearly all-English about banking and nearly all-Hindi about family.
• Register shifts: formal calls drift towards English vocabulary, casual ones towards Hindi, which means a single caller can move along that range mid-call.

Why does code-switching break naive voice stacks?

A voice agent is a pipeline: speech-to-text (ASR), a language model that decides what to say, and text-to-speech (TTS). Each stage can fail on mixed-language input in its own way, and failures compound because later stages trust earlier ones.

• ASR with a single-language model: an English-only model turns Hindi words into phonetically similar English nonsense, while a Hindi-only model mangles English nouns. Either way the downstream model receives corrupted text.
• Language identification done once per call: if the system guesses the language from the first few seconds and locks it, a caller who starts in English and moves to Hindi is mis-transcribed from then on.
• Script confusion: the same sentence can be written in Roman script ('kal subah aana') or Devanagari. If transcripts, knowledge-base entries, and search queries use different scripts, retrieval and matching fail even when the audio was heard correctly.
• Tokenisation of mixed text: language models trained mostly on English handle Romanised Hindi less reliably than English, and weaker still when both appear in one sentence.
• TTS pronunciation: a synthesiser that reads an English brand name with Hindi phonetics, or a Hindi word with English phonetics, sounds wrong to a native listener, and a wrong-sounding voice erodes trust faster than a slightly slow one.

What must a Hinglish voice agent handle?

Roman versus Devanagari transcripts

Decide what you want your records to look like before you evaluate anything. Many Indian teams prefer Roman-script transcripts because their CRM, search, and reviewers' keyboards are Roman-script. Others need Devanagari for Hindi portions. Either is fine, but the agent should be consistent, and its structured outputs (name, date, intent) should not depend on which script the transcript happens to use.

English nouns inside Hindi grammar

Words like 'policy', 'statement', 'slot', or 'refund' must be recognised as English nouns even when surrounded by Hindi verbs, and then mapped to the right intent. A good test is whether the agent understands 'refund ka status batao' and 'mujhe refund chahiye' as different requests, not as the same keyword match.

Numbers, dates, and addresses

Numbers are where Hinglish gets expensive. Callers say phone digits in Hindi ('nau, ek, ek, teen...'), in English, or mixed. They say amounts in lakh and crore, give dates as 'paanch tareekh' or 'five October', and read pin codes digit by digit or in pairs. Addresses add landmarks ('Metro ke paas', 'near the temple') that have no database field. An agent must normalise all of these to one canonical value and read it back for confirmation.

Names and brand words

Indian names have many valid spellings and pronunciations, and ASR often rewrites them as common dictionary words. Brand and product names mixed into Hindi sentences have the same problem. Good systems let you supply a vocabulary of names and terms, and always confirm a captured name or ID rather than trusting a single pass.

What language should the agent reply in?

Understanding mixed speech is only half the job. The agent also needs a language-of-reply policy. There are three sensible options, and the right one depends on your callers and brand.

Whichever you choose, write it down and test it. A frequent bug is an agent that understands mixed speech but replies in stiff, formal Hindi that few callers use day to day, or one that flips language mid-sentence because the model's output is not constrained by the policy.

A 10-point protocol to evaluate any Hinglish voice agent

Run this on your own audio and your own scenarios, not a vendor's demo script. Use at least a few speakers from different cities, record the calls, and score each point as pass, partial, or fail. We deliberately give no benchmark thresholds, because acceptable error depends on your use case. What matters is that you can compare vendors on identical inputs.

1. 01Mid-sentence switch: have a speaker begin a sentence in English and finish it in Hindi, then the reverse. Check the transcript captures both halves.
2. 02English noun, Hindi frame: ask for a refund, reschedule, or balance using Hindi grammar around English nouns. Check the intent is understood.
3. 03Digit strings: read a 10-digit mobile number in Hindi, in English, and mixed. Check the captured number matches exactly, and that the agent reads it back.
4. 04Amounts and dates: say amounts in lakh and thousand, and dates in both 'paanch October' and 'October five' styles. Check normalisation.
5. 05Address with landmarks: give a real address with a landmark and a pin code. Check what is captured, and what the agent asks when something is missing.
6. 06Names: use five names with varied spellings, including ones that resemble common words. Check capture, and whether the agent confirms.
7. 07Script consistency: read the transcripts and the structured fields. Check that Roman and Devanagari output matches your chosen convention.
8. 08Reply policy: check the agent follows your language-of-reply policy across a full call, including when the caller changes language.
9. 09Noise and line quality: repeat a few scenarios on a mobile network, with background noise and speakerphone. Telephone audio is compressed and behaves differently from studio audio.
10. 10Recovery behaviour: say something unintelligible or off-topic. Check the agent asks a short clarifying question in the caller's language rather than looping or guessing.

Two practical tips make the protocol more informative. First, script the scenarios in advance so every vendor hears identical inputs, but have speakers say them naturally instead of reading, because read speech is cleaner than real calls. Second, include at least one speaker who does not mix languages much, since a system tuned only for heavy Hinglish can regress on mostly-Hindi or mostly-English callers.

Common failure modes to listen for

• Language lock: the agent decides the caller is speaking English after a first word and then mis-hears every Hindi phrase.
• Phantom translation: the agent silently translates the caller's words into one language, losing nuance or changing the meaning of a number or negation.
• Digit drift: a correctly heard number is stored with one digit wrong, or read back in a way the caller cannot follow.
• Formal-Hindi drift: replies become so formal that callers stop trusting the call, or the voice reads English brand words with a Hindi accent.
• Knowledge-base miss: the answer exists, but the retrieval step is keyed on Devanagari or English text and cannot match a Romanised Hindi query.
• Repeated apologies: the agent says it did not understand several times rather than escalating to a person.

What Vaami claims, and what it does not

Vaami's published position is that its voice agent handles code-switching, meaning mixed-language speech within a conversation, across Hindi, Hinglish, and several regional Indian languages, and supports 30+ languages overall. Vaami has not published per-language word error rates, so we do not quote one here, and you should not accept a single headline accuracy number from any vendor as a substitute for the test above.

When not to rely on Hinglish alone

• High-stakes legal or medical wording: use fixed, reviewed scripts in a locked language, and have a human confirm anything consequential.
• Regional-language audiences: Hinglish is not a substitute for Tamil, Telugu, Marathi, or Bengali callers. Offer the caller's own language and test it separately.
• Callers with heavy dialect or very noisy lines: plan a human fallback and a clear escalation path.
• Any use where a misheard digit has serious consequences: always read back and confirm identifiers.

For the Hindi side of the same problem, see our page on Vaami's Hindi voice agent, and the glossary entry on code-switching for short definitions of the terms used above.

Sources

• Concepts of matrix language, embedded language, and inter- versus intra-sentential switching: standard terms from the linguistics literature on code-switching (for example Shana Poplack, 'Sometimes I'll start a sentence in Spanish y termino en espanol', Linguistics, 1980; Carol Myers-Scotton, Duelling Languages, 1993).
• Vaami language and capability statements: vaami.ai/india/hindi and vaami.ai/facts.