Hindi, English and both in one sentence

Listen to a real reminder call in North India and you will rarely hear pure Hindi or pure English. You will hear something like: "Haan, EMI toh main next week kar dunga, salary late aayi hai." The borrower moves between languages mid-sentence, uses English for financial words and Hindi for everything else, and expects to be understood.

Most voice systems are built and tested on one language at a time. That is where the trouble starts.

Code-switching is the normal case

For many borrowers, mixing Hindi and English is not an edge case to handle later. It is how they talk about money. "Loan", "EMI", "account", "due date" and "cheque bounce" are usually said in English, even in a sentence that is otherwise Hindi.

A system that assumes one language per call, or decides the language from the first few words, will mishear the rest. A system that answers in formal, textbook Hindi to a borrower speaking everyday Hinglish will sound strange, and the borrower may switch to English simply to be understood. The call should follow the borrower: reply in the language mix they use, and switch when they switch.

Numbers are where it matters most

Getting a casual phrase slightly wrong is tolerable. Getting a number wrong is not.

Numbers in Hindi are hard for speech systems. Each number up to a hundred has its own word, and many sound alike over a phone line. Years are often said the old way, "unnees sau pachaasi" for 1985, rather than digit by digit. Amounts are spoken in lakh and crore, not in thousands and millions, and people mix systems freely: "dedh lakh", "ek lakh bees hazaar", "one twenty".

When a borrower states their year of birth to pass an identity check, or says how much they can pay and when, a misheard number has real consequences. A wrong year fails a genuine borrower. A wrong amount records a promise they never made.

Two habits reduce the damage.

• Read numbers back. After the borrower states a date or amount, the call repeats it in words and asks for confirmation: "Aap baarah tareekh ko pandrah hazaar rupaye bhejenge, sahi hai?" Only a clear yes should be recorded.
• Treat unclear numbers as unknown. If the system is not confident, it should ask again, not pick the most likely value. A flagged gap can be fixed by a person; a confident wrong number is much harder to catch.

Names, accents and phone lines

Indian names come from many languages and are written in many ways. A borrower's name in the loan record may be spelt differently from how it is pronounced. The call should avoid saying a name it is not sure how to pronounce, and should never treat a name it heard as proof of identity.

Accents vary widely across states, and a model tuned on one region's speech can struggle with another's. Phone lines add noise, compression and dropped syllables. Many calls are answered on a two-wheeler, in a shop or beside a television. A system that performs well on clean studio recordings may not hold up on a real line.

Say only approved facts

Language difficulty makes a second risk worse. When a model is unsure what it heard, it is tempting to let it improvise a helpful answer. On a borrower call, that is dangerous. A made-up foreclosure charge or a wrongly stated due date becomes a complaint, and possibly a dispute.

The call should speak only from approved knowledge and the lender's call policy. Amounts and dates should be read from the loan record, never generated. If the borrower asks something the call has no approved answer for, it should say so and arrange for a person to follow up, or hand over with the transcript straight away.

What to test before a voice agent speaks to borrowers

Before any borrower hears an automated call, test it the way borrowers will actually use it.

1. Record real-style test calls in Hindi, English and mixed speech, from speakers across the regions you lend in, over real phone lines, not only in a quiet room.
2. Test numbers on purpose. Years said the old way, amounts in lakh and in thousands, dates said as "das tareekh" and as "tenth". Check what was recorded, not only what was said aloud.
3. Test read-back. Deliberately give a wrong confirmation and check that the call corrects course rather than recording the first value.
4. Test language switches mid-sentence and mid-call, in both directions.
5. Test questions the call should not answer and confirm it declines and offers a person.
6. Test identity checks with wrong and near-miss answers, and confirm that nothing about the account is said before a match.
7. Have native speakers listen to a sample of calls and judge whether the voice sounds natural and respectful, not only whether the words were correct.

Then keep testing after launch. Language on real calls drifts with region, season and product, and the test set should grow with it.

Sonics, the voice part of Inferova, is designed around these habits: Hindi, English and switching between them, read-back before anything is recorded, amounts and dates from the record, and speech recognition and voices that run on the lender's own infrastructure.

If you would like to hear how this sounds on your own call scripts, you can request beta access at inferova.com/demo.