Why Hindi text to speech skips your English words

This is the single most common complaint about Hindi text-to-speech, and almost nobody explains it, so people conclude the tool is broken and try another one that does the same thing.

The failure looks like this

You write a normal Hindi sentence with the English words every real Hindi script contains:

*आज हम WhatsApp के नए feature के बारे में बात करेंगे।*

The audio comes back as "आज हम … के नए … के बारे में बात करेंगे" — grammatical Hindi with two holes in it. Not mangled pronunciation. Absence.

Why absence rather than mispronunciation

A speech model learns a mapping from written symbols to sound, from examples. An Indian-language model is trained on Indian-language speech paired with Indian-language text, which means its training data is overwhelmingly Devanagari, Tamil, Bengali and so on.

Latin characters barely appear in that data. When the model meets a run of symbols it has effectively never seen, it has no learned sound to emit for them — and because these models generate continuously rather than symbol by symbol, the most probable continuation is to carry straight on to the next thing it does recognise. The result is a clean skip rather than a garbled attempt.

Why English text-to-speech has the mirror problem

Run the same sentence through an English engine and the opposite happens. English models have seen Latin script exhaustively, so they read every word — as English. "चलो" written as "chalo" becomes "chah-loh", "paise" rhymes with "raise", and the Hindi is intact but wrong.

Between the two, mispronunciation is easier to catch than deletion, which is the only sense in which it is better.

Fix one: transliterate before generation

Convert the Latin-script words into the target script before the text reaches the model, so it sees only symbols it knows. "WhatsApp" becomes व्हाट्सऐप, and the model pronounces it because now it is Devanagari.

This is what Mayantar does automatically on the Indian-language path. It is a preprocessing step, not a model capability, which is worth knowing because it means you can also do it by hand on any tool that lacks it — writing your English words in Devanagari yourself will fix the skipping on almost any Hindi engine.

It is not perfect. Transliteration has to guess at the Indian pronunciation of a foreign word, and brand names are the hard case: a word can have a settled local pronunciation that is not what a letter-by-letter mapping produces. When a specific brand name comes out wrong, writing it yourself in Devanagari the way you say it is the reliable override.

Fix two: use the code-mixed path instead

If your script is written in Roman letters throughout — "Aaj hum WhatsApp ke naye feature ke baare mein baat karenge" — then it is not Hindi input at all. It is Hinglish, and it needs a path that expects Roman-script Hindi rather than one that expects Devanagari.

This is why Hinglish is a separate language option in Mayantar rather than a setting under Hindi. Choosing Hindi for Roman-script text, or Hinglish for Devanagari, is the most common self-inflicted cause of bad output — and it is one dropdown away from being right.

How to check before you publish

1. Generate one short paragraph that contains your English words, not the whole script.
2. Listen to it against the text, watching specifically for the English words rather than judging overall quality.
3. If a word is missing, write it in Devanagari yourself and regenerate that paragraph.
4. If a word is present but pronounced wrong, also write it in Devanagari — spelled the way you actually say it, not the way it transliterates.

Two minutes of this at the start of a script saves finding a hole in minute nine of a finished video.