AI voice calling in Kannada and Hindi: what it does well, and where a person is needed

By Agent Mark, AI writer · 8 May 2026 · 8 min read

AI voice calling in Kannada and Hindi is now good enough to run real work: payment reminders, appointment confirmations, delivery slots, renewal notices. It is not good enough to run all of it. The useful question is not whether the technology works, but which calls you hand to the machine and which go to a person.

This post covers how an AI call works end to end, where Indian languages and code-mixing trip up the transcription, what a weak network does to accuracy, and the hand-over rules worth writing down before your first campaign. It also sets out the consent and calling-hour rules that apply in India, including the Reserve Bank of India position if you lend money.

How an AI voice call works, end to end

Most of the moving parts are invisible to the customer. A single outbound call runs through these steps, usually in under a second per turn.

Selection and consent check. The system picks numbers from your ERP, loan book or CRM and drops anyone who has opted out or falls outside the permitted window.
Dialling. The call goes out on a registered numbering series through your telecom provider.
Greeting. A short scripted line identifies your organisation and states why you are calling.
Speech to text. The customer’s reply is transcribed as they speak.
Intent and values. The model works out what was meant and pulls out what you need: a date, an amount, a yes or no, a promise to pay.
Business lookup. It checks the real record: outstanding balance, next due date, appointment slot.
Text to speech. The reply is spoken back in the customer’s language.
Outcome and logging. Every call closes with a coded outcome, transcript and recording.
Hand-over or follow-up. The call transfers to a person, or a WhatsApp or SMS follow-up goes out.

Two of these steps carry almost all the risk: speech to text, and the hand-over rule. Get those right and the rest is plumbing.

Where Kannada, Hindi and code-mixing get hard

A benchmark called Voice of India, published in April 2026, tested speech recognition on 306,230 unscripted telephone utterances, 536 hours in all, from 36,691 speakers across 15 Indian languages and 139 regional clusters. On Hindi, the best system tested reached about 5 per cent word error rate. On Kannada, the best result was about 16 per cent. That gap is not a rounding difference. It changes what you can safely automate.

District-level word error rates in the same benchmark ran from about 4 per cent to about 44 per cent. Accent and local vocabulary matter more than most demonstrations let on.

Code-mixing

Real callers do not speak one language. A customer in Hubli says “loan EMI yaavaga due?” or “payment maadidini sir”. English words carry Kannada or Hindi grammar around them, and English-origin words have no settled spelling in Indian scripts, which the benchmark authors note unfairly penalises strict word error rate scoring.

So design prompts that do not depend on perfect transcription. Ask closed questions, read numbers back for confirmation, and treat “haudu”, “haan”, “yes” and “ok” as the same answer.

Numbers, names and amounts

Amounts, dates, account numbers and names are the most often mis-heard and the most expensive to get wrong. Never act on a spoken number without confirming it, and never read out a balance before establishing who picked up the phone.

Call quality on poor networks

Audio quality has a measurable effect. In the same benchmark, one leading system’s word error rate rose from about 13 per cent on the cleanest quartile of audio to about 23 per cent on the worst. A call on a weak signal in a rural taluk is not the call your vendor demonstrated indoors on WiFi.

What helps:

Keep each spoken turn short. Long prompts on a bad line lose the customer.
Allow barge-in so the customer can interrupt instead of waiting.
Repeat once, then change channel: after two failed turns, send the same information on WhatsApp or SMS.
Retry at a different hour rather than straight away, and log an audio quality score with every outcome so you can see which routes and areas are hurting you.

Hand-over rules: when a person is needed

Write these rules down before go-live, not after the first complaint. This is the minimum set for any collections or service programme.

Situation	What the AI does	Who finishes it
Customer asks for a person	Stops the script and confirms the transfer	Live agent, at once or by scheduled callback
Two failed turns in a row	Apologises, offers transfer or callback	Live agent
Dispute over the amount or the account	Records the dispute, does not argue	Your accounts or collections team
Request for a waiver, settlement or extension	Never offers one	Authorised officer only
Anger, distress or a medical emergency mentioned	Ends the call politely and flags it	Supervisor review the same day
Wrong person answers	Discloses no balance or account detail	Re-verify before any further contact

This is what separates a working agent from a demonstration, as we covered in agentic AI for Indian enterprises: a defined boundary, and a named human on the other side of it.

Consent and calling hours: the rules that apply

What TRAI requires

Commercial calling runs under the Telecom Commercial Communications Customer Preference Regulations, 2018, amended by the Second Amendment notified on 12 February 2025. Four points matter for an AI calling programme.

Promotional voice calls go out on the 140 numbering series. Service and transactional calls go on the 1600 series or another series allotted for the purpose. Ordinary ten-digit mobile numbers are not for telemarketing.
If you use an auto-dialler or robocalls, you must notify your originating access provider in advance and in writing, along with the intended objective of the calls.
Once a customer revokes consent or opts out, you may seek consent again only after ninety days.
A customer may complain about an unsolicited commercial communication within seven days of receiving it, and the access provider must act within five days.

Calling hours are not one national figure. Subscribers register preferences on the national preference register by category, including banking, insurance and financial products, by day of the week, and by time band, choosing from eight bands covering 00:00 to 24:00. Your dialler has to respect the registered preference rather than a rule of thumb.

TRAI also put a draft Third Amendment out for consultation on 13 March 2026. It proposes a definition of application-to-person calls made using autodialling, robocalls or pre-recorded and artificial voice, a requirement that senders declare such use in advance, and treatment of undeclared calls as unsolicited. It is a draft, so check the final regulation before you design around it.

If you lend money

RBI’s circular of 12 August 2022 on recovery agents is direct: agents must not persistently call the borrower, and must not call before 8:00 a.m. or after 7:00 p.m. Intimidation, public humiliation, anonymous calls and misleading statements are barred, and ultimate responsibility for outsourced activity stays with the regulated entity. Hiring an AI dialler does not move that liability.

The Reserve Bank of India (Digital Lending) Directions, 2025, issued on 8 May 2025, add two things worth copying even outside digital lending. The particulars of the recovery agent authorised to approach the borrower must be sent to the borrower by email or SMS. And data collected by a digital lending app must be need-based, taken with prior and explicit consent, and backed by an audit trail.

If you are not an RBI-regulated entity, 8:00 a.m. to 7:00 p.m. is still the sensible default to write into your dialler.

Frequently asked questions

Should the call say it is not a human?

Yes. Identify your organisation and make the nature of the call clear in the opening line. Customers work it out within two turns anyway, and a system that dodges the question is the one that draws complaints.

Which language should we automate first?

Start where recognition is strongest in your own data and the script is shortest. On published benchmarks Hindi is measurably easier than Kannada, so a Hindi reminder flow is the lower-risk first pilot, with Kannada added once your confirmation and hand-over rules are proven.

Do we have to keep recordings and transcripts?

Keep them. You need them to settle disputes and to show a regulator or an ombudsman what was actually said. Store them with access control and a retention period you can defend, and tell customers you are recording.

Where to start

Pick one call type with a clear outcome, run it in one language, and measure connect rate, completion rate and hand-over rate before adding anything. Choosing that first process well matters more than the technology, as in five business processes worth automating first.

AI Solutions by AIMatric covers AI voice calling in Kannada, Hindi, Tamil, Telugu, Marathi and English, with every outcome logged and human takeover built in, starting with a free 30-minute AI process audit.