# Builderr starter brief: local dictation

## The job

Build a private dictation tool for Hindi + English speech. A person speaks normally, stops, and gets a clean final transcript that they can paste into another app.

The round is closed, but this is the shortest way to understand and reproduce the task.

## What the agent receives

- Short audio clips containing English, Hindi, and mixed Hindi + English speech.
- The same local machine and offline conditions for every entry.

## What it must return

- The final transcript, keeping Hindi as Hindi instead of translating it away.
- A clean final after speech stops.
- No network calls during scoring.

## What matters

1. Say the words correctly, especially names, numbers, and language switches.
2. Return the final quickly after the speaker stops. The target is about two seconds.
3. Run locally on a normal Mac or Linux machine with commercial-friendly model licenses.

The score is 70 points for the final transcript and 30 points for final-paste latency. Draft text while someone is still speaking is not the score.

## A beginner-friendly way to start

1. Open the [starter repository](https://github.com/builderr-ai/builderr-speech-to-text).
2. Read `GETTING_STARTED.md` and run the included sample clips.
3. Ask an AI coding assistant to explain the `draft()` function and help you change one part at a time.
4. Compare your output with the sample transcripts before trying to improve speed.

You do not need to train a model from scratch. You do need to check that the program runs locally, keeps both languages, and returns the output in the required format.

## Full truth

The [challenge page](https://builderr.ai/speech-to-text) contains the full scoring contract, pinned machine, model-license rule, benchmark, and published results. Read it before submitting. Never commit an API key.