Speech To Text Tool

How it works

The speech model runs in your browser

When you first press the key, your browser downloads a speech recognition model and keeps it. For English it is Moonshine Base, an open model (MIT licence) of about 63 MB. For other languages it is Whisper Base, an open model (Apache 2.0 licence) of about 77 MB. Both run in your browser with WebAssembly, through the open source Transformers.js library. The model files come from the Hugging Face model hub; after that, nothing else is fetched to write your words.

How your speech becomes text

  1. The microphone sound is read in your browser, 30 milliseconds at a time.
  2. A simple loudness check notices when you start talking and when you stop.
  3. When you pause for about three quarters of a second, or after 24 seconds of continuous talk, that stretch of sound is handed to the model.
  4. The model writes the words, with capitals, commas and full stops. Spoken punctuation such as comma or new paragraph is then turned into the marks.
  5. The text is written onto the page and the page is saved in your browser's storage on this device.

What is stored, and where

The current page and up to 50 torn off pages are kept in your browser's local storage on this device. Clearing your browser data removes them. Sound is never stored: each stretch is turned into text and then dropped.

Limits

Accuracy depends on your microphone, the room and how clearly you speak. Names and rare words may need fixing by hand. A slow or older phone takes longer to write each stretch, but keeps up because it works while you talk.