✍🏼 Whisper Dictate

Run once. Hold left Opt+Cmd and speak. It will transcribe and type what you said, so you don't have to.

Great multilingual transcription thanks to Whisper, whisper.cpp!

⚠️ Only tested on macOS.

Installation

1. Make sure you have the rec command from SoX: brew install sox.

2. Install uv.

3. If you will use OpenAI's API:

   1. Set environment variables in your .bashrc or .zshrc:

      export OPENAI_API_KEY=

4. If you prefer to use local whisper.cpp installation (slow unless you are using M1+):

   1. Clone whisper.cpp.

      1. Download the models you want to use. For best results, medium or large (slower).
      2. Follow the instructions to convert models to CoreML. It might take a while!

   2. Install Python bindings for whisper.cpp, with CoreML support (optional) so whisper runs faster.

      pip install --config-settings="--build-option=--accelerate=coreml" whisper-cpp-pybind

   3. Set environment variables in your .bashrc or .zshrc:

      export WHISPER_CPP_ROOT=

      So that WHISPER_CPP_ROOT should point to your whisper.cpp installation. It will get the models from there.

5. Run it with uv run wd and uv will install dependencies for the first time.

6. [Optional] Create an alias in your .bashrc or .zshrc so you can call wd from anywhere:

   alias wd='cd <whisper-dictate path>; uv run wd --openai'

Usage

Just uv run wd, or uv run wd --openai.

Use uv run wd --help for more settings.

If using whisper.cpp, the very first time will take a while to load (whisper.cpp model run through CoreML).

Then hold left Opt+Cmd wherever you are on your system, talk, and lift the keys.

The transcript will be typed and copied to the clipboard.