Whisper Runtime

Turns your speech into text on your own machine, so you can talk to the assistant instead of typing. The audio never leaves the computer.

What you can do

Speak to the assistant in English, or in any other language Whisper knows. English uses a model trained on English alone, which is faster and more accurate; every other language uses the multilingual model.

Getting started

1. Install whisper.cpp so whisper-server is on your PATH (on a Mac: brew install whisper.cpp).
2. Turn on voice in the assistant. The model for your language downloads once, about 490 MB for English and 575 MB for any other language.

Capabilities

Capability	What it does
speech.transcribe	Turns a recording into text with the model for its language. The first call starts the local speech server.
speech.prepare	Downloads the model a language needs, once.
speech.status	Reports which models are downloaded and whether the speech server is running. Reads only.

Requirements

• Desktop (Electron) app 1.63.0 or later, which can send a recording to a local server.
• whisper.cpp's whisper-server on the PATH.