Hinglish AI Calling: Why Code-Switching Is the Real Language Test in India

The first time we listened to a Hindi voice AI demo, the platform's salesperson played a recording of an AI agent saying "Aapka order safal taroop se sweekar kar liya gaya hai." The translation, technically correct, means "your order has been successfully accepted." The salesperson smiled. The Hindi was clean. The pronunciation was textbook. The grammar was flawless.

It was also useless.

Nobody in India says "safal taroop se sweekar kar liya gaya hai" on a phone call. Real customers say "aapka order confirm ho gaya hai." They say "order place ho gaya." They say "transaction successful hai." The AI's beautiful Hindi was the kind of Hindi you hear in a government railway announcement, not the kind you use to actually talk to a customer in Lucknow about their COD order.

This is the Hinglish problem. And it is the problem that separates AI calling platforms that work in production from the ones that work only in conference room demos.

What Code-Switching Actually Is

Code-switching is the technical linguistic term for what every urban Indian already knows: real bilingual speakers don't choose one language and stick with it. They alternate, sometimes word by word, sometimes phrase by phrase, often without conscious thought. Hinglish — the dominant urban Indian communication register — is not an accent or a dialect. It is structured code-switching between Hindi and English, with stable patterns and predictable rules.

A real Hinglish utterance from a customer service call:

"Bhai, order toh place kar diya hai but tracking still pending hai. Payment ho gaya hai UPI se. Kab tak deliver hoga?"

Translated word-for-word: "Bhai (Hindi), order (English) toh (Hindi) place (English) kar (Hindi) diya (Hindi) hai (Hindi) but (English) tracking (English) still (English) pending (English) hai (Hindi). Payment (English) ho gaya hai (Hindi) UPI (English) se (Hindi). Kab tak (Hindi) deliver (English) hoga (Hindi)?"

That single sentence contains six switches between languages, each performing a specific function. The English nouns carry the technical/transactional content (order, payment, UPI, tracking, deliver). The Hindi structure carries the grammatical scaffolding (toh, kar diya hai, ho gaya hai, kab tak hoga). The discourse marker (bhai) is Hindi for emotional register. The connector (but) is English because it's faster than the Hindi equivalent (lekin) in casual speech.

This is not random. It follows patterns. Modern Indian linguistic research has documented at least five distinct types of code-switching that occur reliably in customer service contexts.

Lexical borrowing is the most common — using English nouns inside Hindi grammatical structure. "Order dispatch ho gaya." The grammar is Hindi; the content nouns are English because that is the standard register for transactional vocabulary.

Idiomatic switching uses English idioms inside Hindi flow. "Tension mat lo, I'll handle it." The shift mid-sentence performs a register change — the Hindi part is warm/casual, the English part is action-oriented/professional.

Domain register switch moves to full English for technical or professional content. "EMI bounce hua, NACH mandate re-register karna padega." The financial vocabulary is English (EMI, bounce, NACH, mandate, re-register) because the operational language of Indian banking is English; the verb constructions remain Hindi.

Emotional register switch uses Hindi for warmth, frustration, or empathy and English for formality or distance. A customer angry at a delivery delay will switch into Hindi mid-call. A customer being helped through a complex problem will switch back to English when the resolution becomes formal.

Numeric and brand-name switching keeps numbers, currency amounts, dates, and brand names always in English regardless of surrounding language. "Teen sau pachaas rupaye" in casual conversation, but "₹350" the moment a brand name or transaction enters the picture.

These patterns are stable. A trained Hinglish AI can predict when a customer is about to switch, prepare for it, and respond appropriately. A pure-Hindi AI cannot.

Why "Hindi Support" Is Not Hinglish Support

Most voice AI platforms claim Hindi support. Far fewer support the Hindi that Indian customers actually speak. The technical reasons are worth understanding because they help you ask the right diagnostic questions in vendor evaluations.

The acoustic model — the part of the speech recognition stack that converts sound to phonemes — is trained on training data. A model trained primarily on news broadcasts and audiobook recordings has never heard real telephony audio. Real telephony in India is 8 kHz sampling, lossy compression, mobile network noise, traffic in the background, occasional connection artefacts. A model that performs at 6% Word Error Rate on benchmark Hindi datasets often performs at 25-40% WER on real telephony customer calls because the audio characteristics are completely different.

The language model — the part that decides which words and word-combinations are likely — has to know what "phonepe pe payment kar diya" actually means. A pure Hindi LM will never have seen this construction in training because phonepe, pe, payment, and kar diya don't co-occur in literary Hindi. A bilingual LM trained on Indian customer service transcripts will recognise it instantly.

The pronunciation lexicon — the dictionary that maps written words to their phonetic representations — has to handle words that exist in both scripts. "Order" written in Roman script and "ऑर्डर" written in Devanagari are the same word with the same pronunciation. The lexicon must encode that. Many Hindi STT systems ship with Devanagari-only lexicons, which means English words inside Hindi sentences either fail to recognise or get phonetically distorted.

The intent recognition layer must handle the same intent expressed across multiple linguistic registers. "Cancel karo," "Cancel kar do," "I want to cancel," "Cancel ho sakta hai kya?", and "Cancel please" are five different surface forms of one intent. A pure Hindi system trained only on the first two will fail on the others. A bilingual system trained on real Hinglish transcripts will recognise all five.

The TTS — the text-to-speech voice that the AI uses to respond — must produce natural Hinglish output, not Hindi sentences with English words pronounced in an exaggerated Indian-newsreader register. The voice that says "Aapka order ka delivery date Friday hai" should pronounce "delivery" and "Friday" in casual Indian English, not in formal Sanskritised Hindi-isation. The naturalness of the output is what determines whether customers stay on the call or hang up.

A platform that has solved one or two of these layers but not all five will sound impressive in a curated demo and fail in production. The diagnostic test: ask the vendor to play a real customer recording — not a script-read demo — and listen for whether the AI sounds like a person who actually lives in India.

The WER Benchmark Conversation

Word Error Rate is the standard metric for ASR quality. A model with 10% WER produces, on average, one wrong word in every ten transcribed. Two important caveats apply to interpreting WER numbers in the Indian context.

First, the dataset matters more than the number. A platform reporting 6% WER on the IndicVoices benchmark dataset is reporting performance on clean, studio-recorded, formally-spoken Hindi. The same platform on real telephony customer calls might be at 22% WER. The benchmark and the production reality are different worlds. Always ask: what is your WER on real customer calls from production deployments, not benchmark datasets?

Second, the language register matters. A model trained on broadcast Hindi will report a much lower WER on broadcast-like input than on conversational Hinglish. A platform that quotes its WER without specifying the input register is not giving you usable information. Always ask: what is your WER on the specific register you'll be running for me — customer service Hinglish, BFSI collections register, healthcare patient Hindi?

Realistic WER benchmarks for the Indian market in 2026, based on production data we have visibility into:

• Formal Hindi on clean studio audio: 4-8% WER for well-trained models
• Formal Hindi on telephony audio: 8-14% WER
• Conversational Hinglish on telephony audio: 12-18% WER for top-tier models, 25-40% for poorly-trained ones
• Regional language code-switching (Tanglish, Manglish, Kanglish) on telephony: 15-22% WER for purpose-trained models, 30-50% for adapted Hindi models

The cliff between "well-trained for Hinglish" and "Hindi model adapted for Hinglish" is steep. The cliff is what separates platforms that work in production from platforms that look promising in pilots and degrade as the customer base diversifies.

Hinglish Is Not One Thing — Geography Matters

A Hinglish-aware AI calling platform has to be configurable by region, because what passes for Hinglish in Delhi is materially different from what passes for it in Lucknow.

Delhi/NCR Hinglish. Heavy English content, fast pace, frequent code-switching. The English vocabulary leans aggressive and informal — "scene set hai," "issue create kar raha hai," "drop kar diya." Roman script SMS culture. Customers expect AI agents to keep up with the pace.

Mumbai Hinglish. Marathi substrate creates a slightly different rhythm. Slower switch rate compared to Delhi. More formal Hindi baseline. Specific local lexical items — "scene", "tapori", though these rarely appear in customer service contexts.

Bengaluru tech crowd. English-dominant register, with Hindi switching in only at moments of warmth or escalation. AI calls to Bengaluru customers should default to English-leading Hinglish, switching into more Hindi only if the customer signals comfort with it.

Pune. Younger demographic skews English-dominant; older demographic skews Marathi-Hindi-English with code-switching across all three. Sector matters: D2C calls into Pune skew English-heavy; collections and BFSI calls skew toward Marathi-Hindi.

Tier 2 cities — Lucknow, Jaipur, Indore, Kanpur. More formal Hindi as the baseline. Less English vocabulary, except for specific technical terms (EMI, payment, order, delivery). The customer is comfortable in Hindi but expects you to know the English domain words. Trying to translate "EMI" as "kisht" in this register actually reduces clarity, because customers are used to the English form.

Tier 3 and rural. Hindi-dominant or regional-language-dominant. English appears only as borrowed nouns for unavoidable concepts. The AI must default to formal Hindi or the regional language and use English only for the irreducible vocabulary. A Tier 3 customer in Gorakhpur expecting a Hindi call who gets a Delhi-style Hinglish AI will hang up.

The implication for AI calling deployments: a single Hinglish configuration does not work for all of India. The platform must allow per-campaign or per-customer-segment language configuration. Caller Digital's regional configuration is one of the more underrated reasons production deployments succeed in mixed-geography campaigns.

Writing Scripts for Hinglish — The Practical Guide

The script design problem most teams stumble on: should the AI speak pure Hindi, pure English, or Hinglish? The naive answer is "Hinglish, since that's what customers speak." The better answer is "it depends on the persona, the customer segment, and the call purpose."

Some operating principles that work consistently in our deployments.

Greet in the customer's expected register. A D2C brand calling a Mumbai customer about a beauty product order should open in English-leading Hinglish: "Hi! This is [Brand] ki taraf se calling. Aapka order ke baare mein." A NBFC calling a Lucknow customer about an EMI should open in formal Hindi: "Namaste, main [Lender] se bol raha hoon, aapki kisht ke baare mein."

Use English for transactional vocabulary. Order, payment, EMI, refund, delivery, OTP, UPI, NACH — these are the words customers themselves use, and translating them into Sanskritised Hindi actively reduces comprehension. A customer who hears "aapka rini" instead of "aapka EMI" has to translate twice — once to figure out what rini means, then back to the English term they actually use mentally.

Use Hindi for emotional cues. Empathy, reassurance, urgency, warmth — Hindi carries these registers more naturally for most Indian customers. "Koi tension nahi, hum sambhal lenge" is warmer than "Don't worry, we'll handle it" for a customer in distress. "Bilkul samajh sakta hoon" lands differently than "I completely understand."

Use Hindi sentence structure as the baseline. English nouns and verbs slot into Hindi grammar more naturally than Hindi words slot into English grammar. The default sentence pattern in script writing should be Subject-Object-Verb (Hindi structure) with English nouns and verbs as appropriate. This is what natural Hinglish actually sounds like.

Match the customer's pace. If the customer is responding fast and English-heavy, the AI should accelerate and lean English. If the customer is slower and Hindi-heavy, the AI should slow down and lean Hindi. This requires real-time adaptation in the model — not just a pre-set language configuration.

A worked example. The script for a COD confirmation call for a fashion D2C brand, written in production-quality Hinglish:

"Namaste! Main [Brand Name] se ek AI assistant bol rahi hoon. Aapne ek order place kiya hai — blue kurta, size M, ₹650. Kya aap delivery accept karenge?"

Note the structure: Hindi greeting, brand name in English, AI disclosure in Hinglish (the Hindi "AI assistant" preserves the English term that customers recognise), Hindi grammar carrying the action verbs, English for the product description and price, Hindi closing question with English noun ("delivery").

If the customer responds "Haan ji, kab tak aayega?" — Hindi-leading — the AI continues in Hindi-leading: "Kal evening tak deliver ho jayega. Aapka address confirm karein — [address read back]."

If the customer responds "Yes please, by when will it arrive?" — English-leading — the AI shifts: "It will be delivered by tomorrow evening. Could you confirm your address — [address read back]."

The same script, delivered with real-time register adaptation, produces a natural conversation in either direction. A pure-Hindi or pure-English script produces an unnatural conversation in at least one direction and probably both.

Hinglish by Vertical — Different Rules for Different Use Cases

The Hinglish register is not uniform across industries. Each vertical has its own lexical conventions and tone expectations.

D2C and e-commerce. Most Hinglish-heavy of all verticals. Urban customer base, fast-paced calls, transactional vocabulary entirely in English (order, delivery, payment, refund, return), Hindi grammatical structure. Tone is friendly, slightly casual, customer-service oriented. The abandoned cart recovery script and the post-purchase upsell script both work in this register.

BFSI and collections. More formal Hindi baseline. Technical financial vocabulary stays in English (EMI, NACH, mandate, bounce, payment, account, bureau). Tone is deferential and respectful, especially for collections — the RBI Fair Practices Code requires non-coercive language and the Hinglish register supports this through its Hindi politeness markers. A collections call cannot sound aggressive in compliant Hinglish; the Hindi structure naturally provides face-saving formulations.

Healthcare. More formal Hindi for patient calls, especially for older demographics or Tier 2-3 patients. Medical vocabulary mostly in English (appointment, doctor, lab, report, prescription) but some Hindi terms preferred ("jaanch" for test, "ilaaj" for treatment, "davai" for medicine). Tone is warm, patient-centric, never rushed. Patients in distress need the AI to slow down and switch into more Hindi for warmth.

Real estate. Varies sharply by city. Delhi NCR builder calls are aggressive Hinglish, fast-paced, high-pressure. Tier 2 developer calls are formal Hindi with English for project names and amounts. The script must be configured per developer, per city, sometimes per project — a luxury project sales call uses different Hinglish than an affordable housing project call.

EdTech. Enthusiastic Hinglish — matches the platform's energy. Student conversations are heavily code-switched between Hindi and English in roughly 50/50 proportions. The AI tone should be encouraging, slightly informal, peer-like for student calls and more formal for parent calls.

Insurance. Formal Hindi baseline with mandatory IRDAI English disclosures. The disclosures themselves are English-coded by regulation (policy term, premium amount, exclusion clauses) but the surrounding conversation flows in Hindi. The script writing challenge is integrating the English disclosures into the Hindi flow without sounding artificial.

Hospitality. English-dominant for premium hospitality, Hindi-dominant for value segments. Customer-service tone, warm but professional. The least Hinglish-heavy vertical because hospitality customers expect a polished single-language register.

For voice AI in retail and e-commerce, the calling pattern is heavy Hinglish. For BFSI calling, it leans more formal. The platform configuration has to flex.

Tanglish, Manglish, Kanglish, Banglish — Same Pattern, Different Substrates

Hindi-English is the dominant code-switched register in India, but it is not the only one. Tamil-English (Tanglish), Malayalam-English (Manglish), Kannada-English (Kanglish), Bengali-English (Banglish), Marathi-English, Telugu-English — each follows similar code-switching patterns with the substrate language replaced.

The technical and operational challenges are identical. The same five types of code-switching apply. The same WER cliff between purpose-trained and adapted models applies. The same regional and demographic variation applies. The same script writing principles apply.

A Hinglish-capable AI calling platform should also be capable of:

• Tanglish for Tamil Nadu and Tamil-speaking diaspora — Tamil grammatical structure with English transactional vocabulary
• Telugu-English (Tenglish) for Andhra Pradesh and Telangana
• Kanglish for Karnataka, particularly Bengaluru's local Kannada speakers
• Manglish for Kerala — note that Kerala has high English fluency, so the switching tilts more English-dominant
• Banglish for West Bengal and Bangladeshi diaspora
• Marathi-English for Maharashtra outside Mumbai
• Punjabi-English for Punjab and the Punjabi diaspora
• Gujarati-English for Gujarat

The platforms that handle all of these natively, at production accuracy, on telephony audio, with regional configuration — that is a smaller list than the platforms that claim multi-language Indian support. The diagnostic question for vendor evaluation: "Show me a real customer call recording in Tanglish from a production deployment." If the demo recording is studio-quality Tamil with English words, the platform has not solved the problem.

Measuring Hinglish AI Quality in Your Own Deployment

Once your AI calling programme is live, the metrics that tell you whether the language layer is working are not the same as the metrics that look good in vendor reports.

Real-call WER. Pull 100 random call recordings from your last week of operation. Have a bilingual reviewer transcribe them manually. Compare to the AI's transcription. Calculate the per-call Word Error Rate. Take the average. This is your real WER. Anything above 18% is degrading customer comprehension materially; below 12% is production-grade.

Intent recognition accuracy. For each call where the customer expressed an intent (cancel, reschedule, escalate, confirm), did the AI correctly identify it? Sample 100 calls; calculate per-intent accuracy. Anything below 85% is creating customer friction; above 92% is good.

Fall-through rate. Percentage of calls where the AI said "I didn't understand, can you repeat" more than twice. Above 8% indicates the language layer is failing too often; below 4% is production-grade.

Customer satisfaction by language register. If you run A/B tests with pure-Hindi, pure-English, and Hinglish variants of the same script, which one produces higher CSAT scores? The answer should inform your default register going forward.

Escalation rate to human. A Hinglish-proficient AI should reduce "speak to a human agent" requests. If your escalation rate is high and stable, the AI is failing the language test in ways customers are working around.

Average call duration. A Hinglish-proficient AI should produce shorter calls than one that's struggling, because the customer doesn't have to repeat themselves and the AI doesn't have to ask clarifying questions. Watch this metric over time; rising average duration is a leading indicator of language quality drift.

For AI calling for real estate and other lead qualification deployments, the Hinglish quality gap shows up most starkly in the qualification rate — a Hinglish-fluent AI extracts 30-40% more qualified leads from the same lead pool because it actually understands what customers are saying.

How Caller Digital's Hinglish Stack Works

Briefly, because this is the part where vendor sections get self-serving.

The Caller Digital Hinglish capability is built on three foundations. The acoustic models are trained on Indian telephony audio specifically — 8 kHz sampling, lossy compression, real-world background noise — not on studio-quality benchmark datasets. The language models are bilingual and trained on customer service transcripts from real Indian deployments across D2C, BFSI, healthcare, logistics, and real estate verticals. The TTS layer produces blended voices that pronounce English words in casual Indian English, not in artificial Hindi-isation, and the voices switch register naturally based on the customer's pace and language balance. Production WER on real customer calls runs 8-14% across our deployments depending on the vertical and region.

The deeper context for why this matters for AI calling in India is that the language layer is not a feature — it is the determinant of whether AI calling is a viable channel for your business. Get the Hinglish wrong and every other capability in the platform stops mattering, because customers stop staying on the call.

For the broader voice AI India 2026 picture, the language layer is one of the four foundational capabilities (alongside compliance, integration, and use case fit) that separate production-ready platforms from demo-ready ones. Hinglish is the most under-evaluated of the four because it's the easiest to fake in a demo and the hardest to fake in production.