Your Caller Switched to English. That's Not a Feature — It's Churn

There's a moment on Indian customer support calls that almost nobody measures, and it costs more than any metric on the dashboard.

The caller opens in Hindi. The system doesn't quite follow. So they switch — into halting, effortful English, because that's what the machine seems to want. The call technically completes. It goes in the logs as a success.

What actually happened is that your customer just did work on your behalf, felt slightly small doing it, and learned something about how much you were thinking about them.

The other half of them don't switch. They hang up.

The English-first default, and what it quietly assumes

Most voice AI platforms were built for English and had other languages added later. This isn't a conspiracy; it's just where the training data and the founding markets were. But it produces a specific failure mode: a platform will list Hindi on its language page, and the agent will still be unusable on a real phone call from a real caller in Nagpur.

Because "supports Hindi" is doing an enormous amount of hiding.

Why Indian-language voice AI is actually hard

Five problems, all of which are invisible in a demo and unavoidable in production.

1. Nobody speaks one language. Real Indian speech is code-mixed by default. "Mera payment fail ho gaya, can you check the status?" is not Hindi and it is not English — it's how people actually talk. A system that must choose a language at the start of the call, and stick with it, is fighting the caller. A system that handles the switch mid-sentence, without a beat, is the only one that feels natural.

2. The phone is a hostile channel. Telephone audio is 8kHz, μ-law encoded, mono. Most speech models are trained at 16kHz or higher on clean recordings. That bandwidth loss hits every language — but it hits harder on languages with less training data and denser phoneme inventories. Your agent that scored beautifully on laptop-mic Tamil will meet a caller on a two-bar connection from a moving auto and behave like a different product.

3. "Hindi" isn't a variable, it's a range. Hindi in Lucknow and Hindi in Mumbai are not the same acoustic problem. Neither are Tamil in Chennai and Tamil in Coimbatore. A model tuned on broadcast-standard speech will underperform on the actual distribution of your callers, and word error rate degrades unevenly — which means the failures cluster in a region and you won't see it in the average.

4. The boring things break first. Not the conversation — the numbers. Amounts, dates, order IDs, OTPs, phone numbers, and above all names. An agent that reads back an amount in a way that sounds odd, or mangles a caller's own name, loses trust in a single utterance, no matter how fluent the surrounding sentence was. Read-out conventions differ by language and by region, and getting them wrong is worse than being slightly slower.

5. Prosody carries the meaning. Text-to-speech that gets the words right and the rhythm wrong doesn't sound like an accent. It sounds like a machine. Speech synthesis trained primarily on English intonation applied to Indian-language text produces something callers can hear is fake within about two seconds — which is roughly when they decide whether to trust the call.

None of these are solved by adding a language to a dropdown.

The counterargument worth taking seriously

Here's the honest tension: English is genuinely a unifying force in India.

As Kunal Shah observed in 2019, when a country is divided by castes, languages, religions and cuisines, a neutral foreign identity helps people connect and cooperate — "Pizza for office parties, Cricket for entertainment, English for communication." He's since argued (2025) that "Internet accelerated adoption of English. AI will add jet fuel to it." That's a real argument, and it's why so many Indian businesses default to English support without much soul-searching.

But it conflates two different things: the language of coordination and the language of trust.

English works as neutral ground between strangers who need to transact efficiently. It works much less well the moment a caller is anxious, confused, angry, or trying to explain something that matters — a failed payment, a medical appointment, a delivery that never came. Under stress, people revert. And Shah's own earlier observation (2017) explains why that costs you: "being relatable and having similarities are key levers of trust in any society."

A voice agent speaking your caller's language, in their register, without making them translate their own problem, is the most direct relatability lever available on a phone call. English may be the language of aspiration. It is usually not the language of complaint, and complaint is where most support calls live.

What Telenow does differently

Telenow supports 30+ languages, including Hindi and 10+ other Indian languages — and, critically, it does so by composing the right provider for the language, not by forcing everything through one English-first stack.

Indian languages:

Hindi Marathi Tamil Telugu Kannada Gujarati Punjabi Bengali Malayalam

Plus Indian-English accents — the register a huge share of your callers actually use.

How, specifically:

Indian-language speech via specialists. Sarvam and Smallest.ai are built for Indian languages and Indian phone audio, and Telenow routes to them rather than pretending a globally-trained English model will cope. Global coverage where it fits. ElevenLabs, Amazon Polly, LMNT and Rime for languages where they're strongest, with voice cloning via ElevenLabs if you want your brand's own voice. Swappable, per language. Because Telenow is composable, you're not locked to one vendor's opinion. If a better Tamil STT model ships next quarter, that's a config change and a replay — not a rebuild. Barge-in that survives code-switching. A Rust real-time core with sub-300ms STT partials and configurable interruption handling, so a caller flipping between Hindi and English mid-sentence doesn't derail the turn. Latency you can see, per language. Per-call STT/LLM/TTS/server latency is measured and exposed — so if your Malayalam calls are slower than your Hindi ones, you get a number instead of a hunch.

Language is necessary. It isn't sufficient.

Serving Indian callers well is not only a linguistics problem — it's a telephony and compliance problem, and platforms built for the US quietly leave you to solve it yourself.

Telenow ships the India-shaped parts as part of the product:

DLT registration and regulatory flows, in-app. Not a support ticket, not a PDF someone emails you. Indian carriers as first-class. Run on Exotel, Plivo, Twilio, Vonage or Vobiz — your own account, your own negotiated rates — or a native SIP trunk. Calling windows and DNC enforcement, applied at both upload and dial time, in both directions. Razorpay for payments and prepaid credits, with a hard cut-off at zero and spend caps that mean the bill can't surprise you. WhatsApp as a peer channel, not an afterthought — because for a great many Indian customers, that's the surface they actually live on. One agent, same knowledge, same memory, across phone, web, chat and WhatsApp.

How to test this properly (in any platform, including ours)

Don't trust a language list. Don't trust a demo recorded on a MacBook. Do this instead:

Collect 50 real calls from your actual callers, on actual phones, in actual noise. Include the messy ones — code-mixed, interrupted, background traffic, a bad line. Measure word error rate per language, not on average. The average will hide the language that's failing. Check the read-outs. Amounts, dates, OTPs, names. Especially names. Listen to the TTS with someone who speaks it natively, and ask one question: does this sound like a person, or like a machine reading a script?

If a platform can't survive that test, its language page is marketing.

The point

A caller who has to switch to English to be understood has learned that your company was built for somebody else.

That's a small humiliation, delivered at scale, by a system you paid for. And it shows up in your numbers as an abandonment rate you'll attribute to something else entirely.

Build the agent that meets them where they already are.

Try a live agent in your language → · Read the docs →