What it takes for voice AI to sound local in Hindi and Gujarati

For voice AI to sound local in Hindi and Gujarati, translation is the smallest part. The agent has to understand both languages as they are actually spoken on a phone: full of English words, sometimes transcribed in the other language's script, with prices in lakh and crore and names that do not follow English spelling. It has to reply with grammar that matches its own voice, say numbers the way buyers say them, pronounce local place names as locals do, and follow a caller who switches language halfway through a sentence. Each of these is a separate problem.

Most of it we learned on real calls in Hindi and Gujarati. BlackWolf's agents speak the caller's language, whichever it is: each business's agent is set up in its callers' languages and tried on the business's own questions before a caller hears it. This piece is for anyone choosing an AI voice agent for Indian buyers, or building one. For the basics of how a voice agent works, start with what an AI voice agent is.

Script is not language

A transcript is written in the script the call is set up for. On a Gujarati line, a caller speaking Hindi can come back written in Gujarati letters: every character Gujarati, not one word of it. "Mera naam Rakesh hai, main Surat se hoon" (My name is Rakesh, I am from Surat) can arrive as મેરા નામ રાકેશ હૈ, મેં સુરત સે હૂં, which is the same Hindi sentence in Gujarati letters.

Any agent that decides the caller's language by looking at the script will get this wrong. It sees Gujarati letters, concludes the caller is speaking Gujarati, and keeps answering a Hindi speaker in Gujarati. The opposite happens too: one mumbled word written in Gujarati script can outweigh everything else and drag a Hindi call back into Gujarati.

What survives transliteration is grammar. Hindi and Gujarati share most of the nouns a property buyer uses (parking, possession, price, budget, loan), so counting nouns tells you almost nothing. The small grammar words never cross over:

"Hai" is Hindi whichever alphabet it is written in, and "chhe" is Gujarati. Detection built on these words, listed in all three renderings (Devanagari, Gujarati script and Latin letters), gets the language right where looking at the script fails.

Some words genuinely belong to both. હું (hu) is Gujarati for "I", and it is also how a Gujarati transcript may write the Hindi "hoon" ("am"). One ambiguous word should never change the language of a call. We require a clear margin between the two languages before deciding anything, and a turn that says only "haan" (yes), "ok", a number or a name counts as no evidence at all.

Hindi mixed with English is the normal case

Few buyers speak textbook Hindi on a sales call. They speak Hindi mixed with English, often inside one sentence: "3 BHK ka budget around 90 lakh hai" (the budget for a 3 BHK is around 90 lakh), "possession kab tak milega?" (when will I get possession?), "site visit Sunday ko rakh do" (fix the site visit for Sunday). Gujarati callers do the same: "booking amount shu chhe?" (what is the booking amount?)

Two things follow.

First, English words inside an Indian sentence are correct, not a mistake. The agent should keep project, booking, possession, EMI and site visit in English, because that is how buyers say them. Swapping them for formal Hindi or Gujarati equivalents sounds like a government notice.

Second, the other kind of mixing is a mistake. A Gujarati reply with Hindi words dropped into it sounds wrong to a Gujarati listener even when every fact in it is right. It is also easy to miss: a check that asks "which language is this reply in?" calls it Gujarati, because by majority it is. Mixing needs its own check, which asks a narrower question: is this reply purely the language the caller is using?

Transcripts need the same care. A Hindi transcript in Devanagari is faithful, and slow going for a sales manager who reads Hindi in Latin letters all day. Showing it in the Latin script people actually type ("kal baat karte hain", meaning "let's talk tomorrow", rather than कल बात करते हैं) is transliteration, not translation: the sentence stays exactly what the caller said, in letters the reader moves through quickly.

Gujarati is not Hindi with an accent

An agent asked to speak Gujarati can produce sentences that are Gujarati on the surface and Hindi underneath. A Gujarati listener hears it at once. The tells:

• The question word. Hindi क्या (kya) means "what". Gujarati ક્યાં (kyaan) means "where"; "what" is શું (shu). An agent that reaches for kya while speaking Gujarati is asking the wrong question.
• Postpositions and the verb "to be". Gujarati uses માં (maa, in), થી (thi, from), ને (ne, to) and માટે (maate, for), with છે (chhe, is), છું (chhu, am) and નથી (nathi, is not). Hindi's mein, se, ko, ke liye, hai and nahin (in, from, to, for, is, not) spoken in a Gujarati accent are still Hindi.
• Three genders. Gujarati has masculine, feminine and neuter; Hindi has two. ઘર (ghar, house) is neuter, so "the house is good" is ઘર સારું છે (ghar saaru chhe), never saaro or saari (the masculine and feminine forms). An agent thinking in Hindi has no slot for the neuter.
• Aspirated consonants. Flattening kh to k, gh to g or bh to b sounds like someone reading Gujarati off a page.

Hindi has its own version of this. Verbs agree with the speaker's gender, so an agent with a woman's voice must say "main bata rahi hoon" (I am telling you, in the feminine form), never "main bata raha hoon" (the masculine form). It looks minor on paper and is jarring on a call. The voice and the grammar should come from one setting, so they cannot disagree.

Numbers and money, said aloud

Indian buyers count in lakh and crore, and so should the agent. ₹85 lakh is "pachaasi lakh" (eighty-five lakh); ₹1.2 crore is "ek crore bees lakh" (one crore twenty lakh). An agent that says "eight and a half million rupees" has told the buyer it was not built for them.

Saying the number well is the easy half. Getting the number right is harder, and it matters because budget and price are the core of lead qualification:

• The price must belong to what was asked about. When a project has a 2 BHK and a 3 BHK at different prices, the agent must not quote one for the other. Telling it so is not enough. Checking every price it says against the price list, and correcting a wrong one in the next sentence, is what stops it.
• Numerals come in three scripts. A Gujarati reply writes 45 as ૪૫ and a Hindi one as ४५. Any check that reads numbers out of a reply has to read all three, or it silently checks nothing on Gujarati calls.
• Not every amount is a budget. "The loan is pre-approved for 50 lakh" states a loan, not a budget. So does a rent, an EMI or a salary. A price the caller saw in an ad is not their budget either, and nor is the agent's own quote repeated back as a question. Each of these can end up in a budget field unless it is ruled out.
• Currencies must never blur. For property priced in another currency, such as AED, "10 lakh" said to a Hindi speaker will be heard as rupees. Conversion must be exact, never approximate, and the unit must always be spoken.

Names

Names cause more trouble than their size suggests.

• Not everything is a name. "Haan" (yes) said while thinking is not a name, and neither is "sir", "madam" or "doctor". An agent that stores "Haan" as a name will call the caller "Haan ji" (a polite form of address, as if Haan were their name) for the rest of the conversation.
• Take a name only when it is given. "Main Kiran bol rahi hoon" (this is Kiran speaking, in Hindi), "hu Ronak bolu chhu" (this is Ronak speaking, in Gujarati) and "Pooja here" are introductions. A word the caller happened to say is not.
• Spell names as Indians spell them. त्रिपाठी is Tripathi, not Tripaathee, and सिंह is Singh. Names and ordinary words are romanised to different conventions. Store every name in Latin letters, whatever script it arrived in, or your CRM becomes hard to search.
• Do not open with sir or madam. The agent does not know who picked up. Being called sir can end a call in four words. Better to open without an honorific and use what the caller gives.

Switching mid-sentence

Callers switch languages for reasons of their own. A caller may answer the first question in Gujarati and the fourth in Hindi. A caller asks, in English, "can you speak in Hindi?", and the sentence's own grammar says English while its meaning says Hindi. Naming a language should win outright.

Hearing a switch is not the same as following it. An agent can transcribe the Hindi perfectly and still answer in the default language, and telling it to follow the caller does not fix that reliably. What works is deciding the caller's language on every turn, from the grammar words above, and switching the agent's replies from its very next sentence. We explain the mechanics, and the trade-offs, in how a voice agent remembers and follows a language switch.

Accent and regional pronunciation

• Accent comes from the voice setting, not the brief. If the voice is not set to an Indian language and region, it can fall back to a foreign accent, and no instruction to "sound local" will fix it. The brief shapes the words; the setting picks the voice saying them.
• The accent must survive a switch. When the agent moves from Gujarati into Hindi or English, it should still sound like the same Indian person, not like someone else took the phone.
• Local place names need native spelling, everywhere. Respelling a Gujarati place name in Latin letters does not stop a voice anglicising it. Writing it in Gujarati script does: a Gujarati voice reads it the way it reads any Gujarati word. But the agent has to use that spelling every time, so the Latin version must go from everything it reads, with a check that catches it if it slips through.
• Listening sounds are not interruptions. "Ji", "haan", "hmm" and "achha" (yes, yes, hmm, okay) said while the agent is speaking are agreement. An agent that stops at each one feels broken. Getting this right is partly about voice latency, because the same listening decides when the caller has finished speaking.

Pronunciation and accent are for the ear. Read transcripts for facts, but listen for these.

How to test this yourself

On a trial call with any vendor, try these:

1. Start in Gujarati, then switch to Hindi halfway through a sentence.
2. Ask, in English, for the agent to speak Hindi.
3. Give your budget as "ek crore bees lakh" (one crore twenty lakh) while asking about something else, then see whether it asks for your budget again.
4. Mention a loan amount and check that it is not recorded as your budget.
5. Introduce yourself mid-sentence, and check how the name is spelled in the CRM.
6. Say a local place name and listen to how the agent says it back.
7. Say "haan" and "achha" (yes, okay) while it is talking, and see whether it stops.

For the rest of the checklist, see how to evaluate an AI voice agent vendor.

What to do next

If your buyers speak Hindi or Gujarati, test any agent on a real call in those languages before anything else. For how this works when selling property, see real estate.