Code-switching and Hinglish: Why Most Voice AI Fails in India

"Bhaiya, mera order status kya hai? I placed it last Tuesday."

This sentence has Hindi, English, and implicit Indian context (Tuesday reference without specifying which Tuesday). It is how roughly 500 million Indians speak in everyday commercial contexts.

It is also the sentence that breaks most voice AI platforms.

Code-switching is the biggest unsolved problem in Indian voice AI. And it is the reason many enterprises deploy voice AI, get poor results, and give up before realizing the issue was not voice AI as a technology. It was their platform's inability to handle how Indians actually speak.

Here is what is going on.

What is code-switching?

Code-switching is the practice of alternating between two or more languages within a single conversation or sentence. It happens for many reasons:

Technical terms often have no clean regional language equivalent (account, password, delivery, cashback)

English words feel more neutral for certain contexts (salary, office, boss)

Brand and product names are usually English

Speaker mood or formality shifts

Pure habit from bilingual upbringing

Indian code-switching is especially heavy because:

22 official languages create natural multilingualism

English is a default second language for educated speakers

Hindi is a connector language across regions

Urban Indian culture normalizes mixed speech

Examples:

Hinglish: Hindi + English ("Kya main reschedule kar sakta hun?")

Tanglish: Tamil + English ("Enna delivery charges iruku?")

Tenglish: Telugu + English ("Nenu payment chestanu after 5pm")

Marathi + English ("Maza account kaay balance aahe check kara")

Bengali + English ("Tomar order ta kothay aache?")

Every Indian language has a version of this. And customers expect enterprises to handle it.

Why most voice AI fails on code-switching

Most global voice AI platforms are built on models trained primarily in English, with multilingual support added through translation layers or language-specific models.

When the user speaks pure English, the English model handles it. When the user speaks pure Hindi, the Hindi model handles it. When the user mixes mid-sentence, the platform has to decide which model to route to. The decision is usually wrong. The result is gibberish.

Typical failure modes:

Agent transcribes "kya balance hai" as "kaya balance hi" (mangled Hindi)

Agent responds in pure English when user expects mixed Hindi-English

Agent cannot recognize English product names inside Hindi sentences

Agent keeps switching context with every language change

Agent gets slower (added latency) with each switch

For Indian customers, these failures feel insulting. The technology claims to understand them. It clearly does not.

What good code-switching handling looks like

Strong voice AI handles code-switching as a first-class capability, not an edge case.

Unified understanding

The model processes mixed-language input as a single semantic unit. It does not try to route to separate language models. It understands "kya main reschedule kar sakta hun" as "can I reschedule" with the same clarity as it understands either pure-language version.

Natural response

The agent responds in the register the user is using. If the user is speaking Hinglish, the agent responds in Hinglish, not formal Hindi or pure English. Matching the user's code-switching pattern is key to feeling natural.

Dialect awareness

Within a code-switched conversation, dialect markers still matter. Mumbai Hindi mixed with English sounds different from Delhi Hindi mixed with English. Good voice AI recognizes and adapts.

Context preservation

If a user starts in Tamil, switches to English for a specific query, and switches back to Tamil, the agent preserves context across all switches. It does not lose track of what was discussed when the language changed.

Fast handling

Code-switching cannot add latency. The user experience is real-time conversation. If switches add 500ms of processing per switch, the conversation becomes laggy.

How to test code-switching in vendor evaluations

When evaluating voice AI vendors, stress-test code-switching explicitly.

Ask the demo agent:

A sentence with Hindi and English mixed

A sentence with Tamil and English mixed

A sentence with a brand name in English inside a Hindi question

A sentence that changes language mid-word sometimes

Follow up in a different language than you started

Weak platforms will:

Transcribe incorrectly

Respond in the wrong language

Miss the intent entirely

Introduce delay after the switch

Ask you to repeat yourself

Strong platforms will handle it naturally.

The ARCA approach

At Tenori Labs, ARCA is built for Indian code-switching from the ground up. The architecture treats code-switched speech as the default, not the exception. This is a fundamentally different design choice from platforms that added multilingual support after the fact.

Specifically:

Training data includes heavy code-switched Indian speech

Model architecture does not route between language models; it understands mixed input as a unified semantic stream

Response generation matches the user's code-switching pattern

Dialect awareness is layered on top of code-switching handling

Latency stays under 600ms even through language switches

This is not marketing. It is an architectural choice. And it is why ARCA conversations in Indian contexts feel different from global voice AI platforms.

Why this matters for enterprise outcomes

Code-switching handling directly affects business metrics.

Containment rate: when the agent handles code-switching smoothly, users complete their query with the AI. When it breaks, they transfer to humans. Every broken switch is a containment loss.

Customer satisfaction: users consciously notice and appreciate when an AI handles their natural speech patterns. NPS scores for code-switching-aware voice AI are measurably higher.

Market coverage: enterprises using strong code-switching voice AI can serve customer segments their competitors cannot reach effectively. This is a structural advantage.

The cultural and commercial argument

Code-switching is not a bug. It is how India speaks. An AI that cannot handle it is an AI that cannot serve India.

Western voice AI platforms optimize for pure English. That is the right choice for their markets. It is the wrong choice for Indian enterprises. The right voice AI for India is built with Indian language reality as the primary design constraint.

At Tenori Labs, we made this choice deliberately. We built for India first. International language support came after. This is why ARCA handles 22 Indian languages with code-switching natively, and why our customers report dramatically different conversation quality compared to generic global platforms.

Getting started

If you are evaluating voice AI for Indian customers, make code-switching a top-three evaluation criterion. Test it explicitly. Do not accept "we support Hindi and English" as sufficient. Your customers will mix the two, and you need the agent to handle it.

Talk to us at Tenori Labs if you want to hear ARCA handle real Indian speech patterns. Book a demo and test it with your own code-switched sentences in your preferred Indian languages.