Why Indian languages are still hard for AI voice

AI voices in English now sound close to human. Switch to Gujarati, Tamil or Bengali and the gap shows up within a sentence or two. AI text to speech in Indian languages has improved quickly, but it is still harder than it looks, and the reasons are more interesting than "not enough data".

I build Bolo Media, which makes video ads with natural Gujarati and Hindi voices, so this problem is never far from my desk. Gujarati will be my main example here because it is my language, but almost everything below applies across India's languages.

India is not one language problem

India has 22 languages in the Eighth Schedule of the Constitution and hundreds more spoken at home. Many have their own script. Hindi and Marathi share Devanagari. Gujarati has its own script, close to Devanagari but without the top line. Tamil, Telugu, Kannada, Malayalam, Bengali, Odia and Punjabi each have theirs.

For an AI voice, this means there is no single "Indian" model to tune. Each language has its own sounds, rhythm and spelling habits. A voice that does well in Hindi can still stumble in Gujarati, even though the two are close cousins.

Why AI text to speech in Indian languages still slips

Here are the patterns that come up again and again. None of them is a secret. They are just hard.

The vowel you write but do not say

Indian scripts are built so that each consonant carries a short "a" sound by default. In speech, that sound is often dropped, especially at the end of a word. Linguists call this schwa deletion. In Hindi, कमल is written with three letters that could be read "ka-ma-la", but people say "kamal". Gujarati does the same: વાત is spoken "vaat", not "vaata".

The rules for when to drop the vowel are not fully written in the spelling. Native speakers just know. A model that reads too literally sounds robotic, like a child sounding out letters.

Code-mixing is the normal way people speak

Real Indians do not speak "pure" Gujarati or Hindi in an ad. They say "offer", "free delivery", "WhatsApp karo", "best quality". The English word may be written in Gujarati script, in Roman letters, or both in the same line. A voice has to decide whether "sale" should sound English or Gujarati-accented, and the right answer depends on the speaker and the audience.

Many models are trained on cleaner, more formal text than what people actually say. So they handle news-reader sentences better than they handle a shop's own words.

Numbers, money and dates

Indian numbers are grouped in lakhs and crores, not millions. "1,50,000" should be read as "ek lakh pachaas hazaar" in Hindi, not "one hundred fifty thousand". Prices, phone numbers, dates and discounts each have their own spoken form, and that form changes by language. A wrong reading of a price in an ad is not a small mistake. It is the one detail the viewer remembers.

Names, places and brands

Shop names, town names, family names and product brands are rarely in any dictionary. A textile shop named after a family deity, or a sweet shop named after its founder, has to sound the way the owner says it, not the way a model guesses. This is where AI voice fails most visibly, because a business owner notices immediately when their own name sounds wrong.

Sounds English does not have

Indian languages have contrasts many models do not hear clearly: aspirated and unaspirated consonants, dental and retroflex sounds, nasal vowels. Gujarati has ળ, a retroflex "l" that does not exist in English. Get these slightly wrong and the voice sounds foreign, even if every word is technically right.

Data is part of it, but not all of it

It is true that Indian languages have less clean, labelled speech data than English. But more data alone does not fix everything above. Schwa deletion, code-mixing and number reading are about rules and context, not only volume. A model needs to know who is speaking, to whom, and in what register.

Tone matters too. A wedding-wear shop and a tractor-parts dealer should not sound the same. Indian ads often use a warm, slightly conversational tone, closer to talking to a neighbour than to reading a script. Most voices default to something flatter.

What actually helps

From the builder's side, these are the things I believe make the biggest difference today.

• Write for the ear, not the eye. A script that reads well on paper may sound stiff aloud. Short sentences and natural phrasing help any model.
• Spell tricky words the way they are said. Sometimes a small change in spelling fixes a pronunciation better than any setting.
• Handle numbers before the voice sees them. Convert prices and dates to the words you want spoken.
• Listen to every output. A native speaker catches in five seconds what no automated check will.
• Test with real words. Shop names, local places and mixed-language phrases, not textbook sentences.

Bolo has a hands-on post on exactly this for Gujarati: Gujarati words that break AI voice. It goes into specific words and fixes in more depth than I can here.

Why I care so much about getting it right

For a small business, the voice is the ad. If it sounds off, people scroll past, and the owner feels embarrassed rather than proud. That is a real cost, even if it never shows up on an invoice.

I have written about why the vernacular internet is not a translation problem, and about building in the language I grew up speaking. This post is the technical side of the same belief: Indian languages deserve voices that sound like the people who speak them.

The tools are getting better every few months, and I am genuinely optimistic. But "good enough in English" is not the bar. The bar is whether an aunty in Surat or a farmer near Rajkot hears it and thinks, yes, that is how we talk. We are not fully there yet, and I think saying so plainly is the first step to getting there.