Hinglish podcast clips

The Challenge

The show is two people talking the way people actually talk in Delhi — a Hindi sentence with an English clause dropped in the middle, a punchline that starts in one language and lands in the other. Every transcription tool they tried treated that as a problem to be solved by picking a side. Set it to Hindi and the English came back as nonsense syllables. Set it to English and the Hindi vanished into guesswork.

So captions were done by hand. Roughly forty minutes per clip, three to five clips an episode, every week. The editor doing it was the same person cutting the episode, which meant captions were always the last thing standing between a finished cut and a post.

Our Strategy

The audio goes through a model that was trained on code-mixed Indian speech rather than one trained on English and retrofitted. The line stays whole: Hindi words come back in Hindi, English words in English, and the sentence keeps the order it was said in.

Where the transcript is wrong, it gets fixed a word at a time. Dragging one word on the timeline moves that word only — the words either side hold their timing, so correcting a name does not push the rest of the line out of sync with the audio.

One template is set for the channel and reused. The same caption look lands on every clip without anyone opening a style panel, which is what makes the back catalogue worth recaptioning.

The Results

Captions stopped being the last hour of the edit.

The same clip now comes back transcribed in one pass, with the code-switching intact, and the editor's job is a read-through rather than a retype. Three output formats come from that single transcript: Romanized for the caption burn, Devanagari for the Hindi-first audience, and English for the clips that travel.