Voice AI

India

Callers reply in a language you can't cover? Start your AI voice bot from the call log

By Shivanath DevinarayananPublished 7 min read
Written forChief Collections Officer at an Indian NBFC who answers for every call the team makes“Half my telecallers speak two of the eight languages our borrowers speak.”

The short answer

Which language should an Indian service or lending team start an AI voice bot with?

Start with the language pair that carries the most calls where your team's cover is thinnest, on one process, picked from your own call log and not from a vendor's language list. Test 50 of your own recordings per language before the bot dials. At Mindcat, we call this spec, shadow, gate.

A call log tallied by language with one pair circled and a headset beside it

Takeaways

  1. 01Count calls by the language the caller answered in. The pair with the most calls and thinnest cover goes first.
  2. 02Sarvam's recognition hears 22 Indian languages and English, its voice agents answer in 11 plus English (read 3 November 2026). Heard is not answered.
  3. 03A language passes only when a reviewer who speaks it marks 50 of your own recordings: heard, and spoken naturally.
  4. 04Consent and calling hours belong to your call flow. Sarvam's own guide says to capture consent in the call flow itself.
  5. 05Platform per-minute prices are the smaller question. The first one is whether the language is covered at all.

If callers answer in a language your team does not speak, this is for you. It is not for teams with no recorded calls to count from, and it is not for lenders who want a dialler to call more numbers.

Which language do your callers actually answer in?

Open last month's call log and count. Not the language you called in, the language the person on the other end answered in.

Most teams do not have that number. They have a campaign language and a feeling that "Tamil is hard to staff at night." The count turns the feeling into a pair. Say Hindi to Marathi on weekday evenings, or Malayalam to English at month-end. One pair, one process, one window of the day.

Our rule, so you can argue with it. Pick the first language from your own call log, not from a vendor's language list. Find the language pair that carries the most calls where your cover is thinnest: nights, peaks, a branch with nobody who speaks it. Start with that pair on one process that reads fixed facts, such as a due-date reminder, an appointment confirmation or a status question. Everything else stays with a person until a test on your own recordings says otherwise.

If your calls land in a loan or service record, see how our India voice agents are set up, and which Sarvam use cases teams start with.

Why does hearing a language not mean answering in it?

Because they are two different models with two different lists. This is the receipt for the rule, and it comes from Sarvam's own pages, read on 3 November 2026.

JobModelLanguages, per Sarvam
Hear the caller (speech recognition)Saaras v422 Indian languages and English, with code-mixed audio
Answer in a voice (speech synthesis)Bulbul v3 and v4Listed on the Voice Agents page as 11 Indian languages plus English
Run the whole call (Voice Agents)Saaras, an LLM and Bulbul together11 Indian languages plus English: Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese

Sources: Sarvam, voice agent models and why Voice Agents.

Read the gap. A caller who answers in a language that is in the 22 and not in the 11 can be transcribed, and the agent still has no voice to answer in. Our Sarvam page, checked on 10 October 2026, said ten spoken languages, and Sarvam's page now lists eleven with Assamese. The number on Sarvam's page on the day is the one that counts, so we date ours. Urdu is the example we state on that page: Saaras can transcribe it, Bulbul cannot speak it.

That is a design input, not a fault. A call that is heard but cannot be answered in goes to a person who speaks the language, or the agent hands over with the transcript. Your call log tells you how many calls that is before you build anything.

Two boxes: hears the caller in 22 Indian languages and English, answers in a voice in 11 Indian languages plus English; a gap goes to a person
Heard is not answered. Sarvam's pages, read 3 November 2026.

How do you pick the first language from your call log?

An afternoon, with last month's log. Five steps.

  1. 01Export a month of calls with the language the caller answered in. If the log has none, tag a sample of 200 by listening.
  2. 02Count calls per language, then per language pair if your team switches mid-call.
  3. 03Mark the hours and branches where nobody on the rota speaks that language.
  4. 04Cross out every language that is not in the voice list for your platform, and write down which of your own languages are out. Those stay with a person.
  5. 05Take the pair with the most calls left, and one process inside it: a reminder, a confirmation or a status question.

Step four is the one teams skip. A platform's language page is where you cross candidates out, not where you pick. The pair you start with is the biggest gap that survives it.

Five steps: count calls by language, find the thinnest cover, cross out unspoken languages, pick one process, test 50 recordings
Picking the first language, in order.

What does the shadow run on 50 recordings show?

Before the bot dials anyone, it listens. In shadow mode the agent transcribes 50 of your own recorded calls per language and drafts the reply it would have given. A reviewer who speaks the language marks each one on two questions: did it hear what was said, and is the reply how a person would say it. A language passes on the reviewer's mark and nobody else's. The pass mark, for example how many of the 50 must be marked both heard and natural, is written into the spec before the run and signed by your compliance head, not set by us after the fact.

Voice vendors give the same advice. Sicada, a voice vendor, writes on its blog dated 4 July 2026: "pull 50 anonymised recordings from your actual contact centre and run them through any ASR model before committing; vendor benchmarks use clean studio audio; your calls do not." Convo360, another vendor, in a page updated on 19 September 2026, says to test every script with a native speaker before it goes live, "especially for numbers, dates and product names."

Numbers and dates are where a polite reply goes wrong. An amount said in lakh, a due date said as "the fifth", a loan number read out in two languages. The reviewer listens for exactly those.

This is the "shadow" in our method, spec, shadow, gate: the process is written down, the agent runs beside a person without acting, and a human gate sits on anything that writes to your record.

Four steps: 50 recordings per language, agent drafts a reply, a reviewer who speaks it marks, pass or stays with a person
The shadow run before any call.

Two more things go in the spec before any language is switched on. First, what lands on the record: our India agents page states that the outcome, the promise-to-pay date, the fields and the recording link are written back to the loan in Salesforce, and a named reviewer checks those writes. Second, who stops it: a named person on that day's rota who can pause the dialler, with a cover when they are out.

What does a voice platform cost, and what is left over?

Sarvam's pricing page, read on 3 November 2026, lists these per-minute rates for its Voice Agents. They are Sarvam's list prices in rupees, they exclude GST, and Sarvam says telephony is billed separately.

PlanRate per minuteMinimum subscription
Pay as you goRs 3.50None
ProRs 3.25Rs 9,999
GrowthRs 3.15Rs 49,999
BusinessRs 3.00Rs 1,99,999
EnterpriseCustomFrom 2 lakh minutes a month

Source: Sarvam, Voice Agents pricing. A call is rounded up to a 60-second pulse, so a 45-second call counts as a minute.

Your own cost per call is minutes per call times the rate, plus telephony, plus the minutes a person spends on every call that is handed over. Use your own log for each. An example only: a 90-second call is two billed minutes, so Rs 7.00 on pay as you go before telephony.

Notice what the table does not decide. It does not say which language is covered, which calls a person keeps, or what is written to your record. Those are the first-language question and the call flow. They are why we do not start from a price. On your own number you hold the account with Sarvam or your voice provider, and our build is quoted separately after the call. Mindcat is not affiliated with Sarvam.

You do, in your call flow. Sarvam's own contact-centre guide lists "Call recording (with consent)" and says that if you are recording for quality checks, "capture consent in the call flow itself." The platform runs the call. The notice, the recording rule and the hours are yours.

So the spec for the first process names them before the first call: the window the dialler may call in, the line the caller hears about recording, how a caller opts out, and who is the named person when they do. For lenders in India, RBI's recovery Directions add paragraphs on the log, the recording and the hours from 1 January 2027, and we read them in which collection calls keep a person. Your compliance team decides how any rule applies to your licence.

The platform runs speech recognition, voice and billing; your call flow owns consent, calling hours, opt-out and a named person
Who owns what on the call.

When are we the wrong answer?

When you have no recorded calls to count from, or you want a dialler that places more calls. Without the log, the first step is to start recording with consent and count for a month. Our First Agent in Production offer starts with one process in one language pair, a shadow run on your recordings and a reviewer on the exceptions. The first look is free.

If you want the written read, send us the call your team takes in the wrong language and the system it lands in. We reply in writing within 24 hours.

Filed under

Voice AIai voice bot indiamultilingual voice bot indiaindian language voice agentvoice bot for collections india
Useful to someone on your team?

FAQ

Questions teams ask next

Which language should an AI voice bot start with?
The language pair that carries the most calls where your team's cover is thinnest, on one process such as a due-date reminder or an appointment confirmation. Count last month's calls by the language the caller answered in. Do not start from the vendor's language list.
Does an Indian-language voice bot understand and speak the same languages?
Not always. Sarvam's page for its speech recognition model, Saaras v4, says it covers 22 Indian languages and English. Its Voice Agents page lists 11 Indian languages plus English for the agent that answers. Both pages were read on 3 November 2026. A language in the first list and not the second can be heard, and the caller still needs a person or another language back.
How many recordings should I test before an AI voice bot calls my customers?
We use 50 of your own anonymised recordings per language. Sicada, a voice vendor, advises the same on its blog dated 4 July 2026: pull 50 anonymised recordings from your own contact centre, because vendor benchmarks use clean studio audio. A reviewer who speaks the language marks each one.
How much does Sarvam's voice agent platform cost per minute?
Sarvam's pricing page, read on 3 November 2026, lists Rs 3.50 a minute on pay as you go, Rs 3.25 on Pro (Rs 9,999 minimum a month), Rs 3.15 on Growth (Rs 49,999) and Rs 3.00 on Business (Rs 1,99,999). Prices exclude GST. Telephony is billed separately. These are Sarvam's list prices. Our work is quoted separately after the call.
Who is responsible for consent on a recorded AI call?
The lender or business that places the call. Sarvam's contact-centre guide says to capture consent in the call flow itself when you record for quality checks. Your compliance team decides what the notice says and how long recordings are kept.

Sources

  1. 01Sarvam: Voice agent models, Saaras ASR and Bulbul TTS (read 3 November 2026)Accessed 03 Nov 2026
  2. 02Sarvam: Why Sarvam Voice Agents, build vs buy (read 3 November 2026)Accessed 03 Nov 2026
  3. 03Sarvam: Voice Agents pricing, per-minute rates in INR (read 3 November 2026)Accessed 03 Nov 2026
  4. 04Sarvam: IVR and contact centre guide (read 3 November 2026)Accessed 03 Nov 2026
  5. 05Sicada: Multilingual voice bots, best practices for India, SEA and MENA (4 July 2026)Accessed 03 Nov 2026
  6. 06Convo360: Regional-language voice bots (updated 19 September 2026)Accessed 03 Nov 2026
Shivanath Devinarayanan, founder of Mindcat

Written by

Shivanath Devinarayanan

Founder, Mindcat Consulting · Salesforce MVP Hall of Fame

Builds and runs AI agents in production. Reads every brief that comes in and signs the work that goes out.

About Shivanath

First Agent in Production · the written assessment

We work in real estate and right now, the AI pilot never left the sandbox.

Industry
Problem

Brief: Real estate. Leads from portals and WhatsApp. Follow-up has no owner. Stalled AI or Salesforce pilot. Need a finish, cut, or rebuild call. From: /blog/ai-voice-bot-first-language-india.