Pick the script before the words
Hindi is written in Devanagari and Arabic in Arabic script, and text-to-speech voices are trained on those scripts. Plenty of people type both in Latin letters online, as Hinglish and Arabizi, but that is not what a voice reads best. Hyzgen writes Hindi and Arabic scripts in their own script, and a line typed in Latin letters is transliterated before the voice reads it.
Hindi is the forgiving one: the Hindi voices we use read Hinglish acceptably either way. Our firmest number comes from Urdu, which Hindi speakers largely understand when spoken: an Urdu voice reading a line in Latin letters scored 1 out of 10, with half its words missing, and the same line in Urdu script was fluent.
Choose the variety as well as the language. Hyzgen's Arabic voices are Saudi (Zariyah and Hamed) and filmed takes are directed as natural Gulf Arabic; the Hindi voices are Swara and Madhur, from India. Spoken Arabic differs a lot from country to country, so an ad for Egypt or Morocco needs lines written the way people there speak.
English words stay in English letters
Hindi and Arabic ads borrow English product words all the time: app, order, delivery, offer, WhatsApp. Written in Devanagari or Arabic letters, a voice can read them as native words and say them wrong. Write them in English letters inside the line instead: अपना menu खुद बनाया. The rest of the line stays in its own script.
Mixed lines need care further down the line too. When a take is transcribed to check it against the script, tell the transcriber which script the line is in. An Urdu line that was half English words came back transcribed in Devanagari, and the check, comparing two different scripts, scored a good take at 50 percent and ordered a retake nobody needed.
Captions: shaping and direction
Arabic letters join and change shape with their neighbours, and the line runs right to left. Devanagari stacks vowel signs and joins consonants. Both need a shaping engine such as HarfBuzz before they are drawn, and some widely used caption renderers (libass in many static ffmpeg builds) ship without it, which turns the text into loose or misordered letters.
Hyzgen draws Arabic captions as shaped outlines in Noto Sans Arabic, so they look the same in every player. Hindi captions are romanised, which is also how many Hindi creators caption; Devanagari captions need a renderer with shaping, so check yours by rendering a real line, not by trusting that the font is installed.
A mixed line reads right to left with its English words left to right inside it, the way a phone shows it. And keep each caption cue inside one shot: a cue that runs across a cut mixes two people's lines.
Numbers in Arabic text
After Arabic letters, digits follow the sentence's right-to-left order, and neither a space nor a hyphen holds two groups of digits together, so a phone number written as two groups is shown with the groups swapped. Put a left-to-right mark (U+200E) just before the number. It works in captions, in WhatsApp and in a post's text. Hindi runs left to right and has no such problem.
Make the scene fit the market
Image and video models default to Western scenes. Hyzgen sets every ad scene in the brand's own market, read from the places the brand names, its currency, its domain and the ad's language, and keeps alcohol out of scenes for Muslim-majority markets.
Saying so is not always enough. Told "no alcohol", an image model still put red drinks in wine glasses on a wedding table. What worked was naming what is on the table: juice, tea, lassi or soft drinks in plain glasses. The same goes for people and places: name the city, the clothes and the dishes, and the model stops reaching for its defaults.
Fit the words to the clip
Filmed clips have a fixed length, about 10 seconds a clip with Gemini Omni. We have firm numbers for two languages: 14 to 17 words per 10 seconds in Urdu, 16 to 20 in English. For Hindi and Arabic, start near the Urdu figure and listen for the sign of a line that is too long: the last word cut off. Too few words is a problem as well, because the video model fills a silence with words of its own.
A checklist before you publish
- The voice line is in Devanagari or Arabic script, in the variety your buyers speak.
- English product words are in English letters inside the line.
- Captions are shaped, or cleanly romanised, never loose letters.
- Every phone number in Arabic text has a left-to-right mark before it.
- Scenes, people, food and drinks belong to the market.
- You listened to every take, last words included.