Skip to content
vaak

Live demo

Two modes. Both take a microphone recording or an uploaded file.

  • In your browser: downloads onnx-community/whisper-tiny (about 40 MB) once, then runs it locally with WebGPU or WASM. Nothing is uploaded. This is the smallest Whisper model, so expect rough output, especially for Hindi. It is here to show the pipeline, not to represent vaak’s accuracy.
  • Your own vaak server: start vaak serve, enter the base URL, and the page calls POST /v1/transcribe. This runs whichever backend you loaded.

Notes

Section titled “Notes”
  • The in-browser model is served from the Hugging Face Hub on first use. The ONNX Runtime WASM binary is fetched from a CDN by transformers.js on first use. After that the browser cache serves both, so the demo works offline.
  • Microphone capture needs a secure context (https:// or localhost).
  • The server mode sends a cross-origin request. Run vaak serve with CORS enabled for this site’s origin, or open the docs locally with npm run dev on the same machine.
  • Nothing from this page is sent to any third party.