The problem, in one clip
Run the same sentence through most caption tools twice, in two videos with different amounts of English, and you get two different-looking results. Neither is wrong. They just do not match each other, and a feed of mismatched captions looks careless.
What you usually get
यार बहुत अच्छा
yaar bahut accha
Same person, two clips.
What you get here
yaar bahut accha
yeh actually works
Every clip, the same.
Spelled the way people write it
Converting Devanagari letter by letter gives you words nobody would type. Hindi drops vowels that the script still writes, so a literal conversion reads like Sanskrit. These are handled properly:
| Devanagari | Literal conversion | What we write |
|---|---|---|
| इतना | itanaa | itna |
| उसका | usakaa | uska |
| काफ़ी | kaafee | kaafi |
| अच्छा | achchhaa | accha |
| करके | karake | karke |
Questions
What is a Hinglish caption?
A caption for speech that mixes Hindi and English in the same sentence, which is how most Indian creators actually talk. The difficulty is not the words but the script: the Hindi parts can be written in Devanagari or spelled out in Roman letters, and most tools switch between the two without warning.
Why do captions change script between my videos?
Speech recognition detects one language per clip. If a video is mostly English it returns Roman script throughout; if it leans Hindi it returns Devanagari for the Hindi words and Roman for the English ones. Same speaker, different clip, different-looking captions.
How does this fix it?
You choose the script once and it holds for every clip. Choosing Roman converts any Devanagari in the transcript into Roman spelling, so the captions look consistent whatever the mix of languages happened to be.
Does the conversion spell words the way people actually write them?
That is what it is tuned for. A direct letter-by-letter conversion produces things like "itanaa" and "usakaa"; Hindi drops those vowels in speech, so this writes "itna" and "uska". Around sixty of the most common words are spelled the way creators write them rather than the way a converter would.
Which other Indian languages work?
Speech recognition covers 99 languages including Tamil, Telugu, Bengali, Marathi, Punjabi and Gujarati. The script conversion is currently Devanagari to Roman only, so Hindi and Marathi benefit; other scripts come back as detected.