Bol Ke Bharo: speak Hindi, fill an English PDF form

The form is in English, and mummy doesn't know English. Bol Ke Bharo is a small voice app for that case. You speak Hindi, and an English form gets filled in for you.

It's the first entry in the Build This series, and it fits in a weekend.

How the pipeline works

There are three steps, and each one does a single job.

First, speech to text turns the spoken audio into text. The person talks in Hindi, the app gets words back.

Second, an LLM reads that text and works out the English value for every field on the form. This is the step that does the translating. It doesn't just transcribe, it decides what goes into each box. Someone saying their details in Hindi gets an English answer per field.

Third, pdf-lib takes those values and fills the PDF. The output is the same form, now complete.

Keeping the stages separate helps. If a field comes out wrong, you can look at the text, then at the extracted values, then at the PDF, and see where it went off.

The hard part is names

Most fields are forgiving. Names are not.

Say the app writes Sunita as Suneeta. Both are reasonable ways to spell the same sound in English. But the spelling has to match the ID, and the ID has one spelling. The wrong one won't match.

An LLM will happily pick a spelling that looks plausible. That's the problem. Plausible isn't the same as correct, and nothing in the audio tells the model which spelling is on the ID.

So the rule for this project is simple. Never guess a name. If the app isn't sure of a spelling, the answer is to ask the user rather than choose for them.

A realistic scope

The plan is Saturday for one form, Sunday for a review screen.

One form on Saturday keeps the work small. You wire up speech to text, the extraction step and the PDF fill for a single document, and you don't try to cover every form at once.

Then Sunday goes to the review screen, which gives you a place to look over the result before you call the weekend done.

What to try

Pick one form that someone you know struggles with, and build the three steps for just that form. Treat names as something you never guess.

The next reel in the series covers a six minute voice note in three lines. Comment BUILD on the reel and the blueprint gets sent to you. And if you want to leave a thought, tell us which form you'd build this for first.