Skip to content
LedgeurLedgeur
8 June 2026 · 6 min read

How Whisper Runs in Your Browser (No Server Needed)

A plain-English explanation of how OpenAI's Whisper model runs entirely in the browser with WebGPU and WebAssembly — and why that's a big deal.

It used to be that speech-to-text required a server with a GPU. Now it can run in a browser tab. Here's how that became possible and what it unlocks.

The pieces

  • Whisper — OpenAI's open speech-recognition model, available in small, fast variants
  • Transformers.js — runs Hugging Face models in JavaScript
  • WebGPU / WebAssembly — let the browser do heavy maths efficiently

What happens when you record

The browser captures your audio, resamples it to 16 kHz, and feeds it to a Whisper model that was downloaded once from a CDN and cached. The model produces text — all on your device. No audio is ever uploaded.

Why it's a big deal

It removes the server from the privacy equation entirely, makes the tool free to run at scale, and works offline after the first model load. Ledgeur is built on exactly this stack.

Try it on a meeting you have already recorded.

Drag a recording into Ledgeur and get a transcript with the speakers separated — in your browser, with nothing uploaded. Free, permanently, and no account needed.

Open LedgeurSee pricing