# openvox | In-browser AI by Shubham Prakash

> Talk to GitHub. Name a profile or repository and it introduces it out loud, then answers spoken questions. Speech, retrieval and the model run in the browser.

Page: https://shubhamcodess.github.io/builds/openvox/

## openvox

Talk to GitHub. Name a profile or repository and it introduces it out loud, then answers spoken questions. Speech, retrieval and the model run in the browser.

- Page: https://shubhamcodess.github.io/builds/openvox/
- Kind: In-browser AI, 2026
- Stack: React, TypeScript, WebGPU, WebLLM
- live: https://shubhamcodess.github.io/openvox/
- code: https://github.com/shubhamcodess/openvox

### Overview

Type a GitHub username or repository. It **introduces it out loud**, then answers spoken questions in a natural voice. There is no backend and no API key, and **audio never leaves the browser**.

Measured:

- 0: servers. Audio never leaves the browser
- 9/10: on the built-in 10-question eval, single run

### What it does

- A spoken intro that uses only numbers computed in code, never numbers a model made up.
- Voice or text questions, answered from profile facts and READMEs, with the sources shown.
- Says plainly when GitHub cannot show something, instead of guessing.
- Start talking and it stops. It also waits through thinking pauses.

### How it is built

- The pipeline runs end to end in the tab: voice detection, speech to text, question routing, BM25 retrieval over README sections, a 1.7B model on WebGPU, then streamed speech.
- Speech recognition runs on WASM and the model on WebGPU, in separate workers, so they do not compete for the GPU.
- A speech normalizer turns links, commands and acronyms into text that can be spoken.
- The built-in eval mixes fact lookups with trap questions, and was used to pick the model.

### Honest limits

- Needs a WebGPU browser, such as current Chrome or Edge.
- About 1 GB of model to download on first load.
- Public data only, and 60 GitHub requests an hour.

By Shubham Prakash: https://shubhamcodess.github.io/
