push2whisper

A push-to-talk speech transcription tool powered by whisper.cpp. Press a hotkey, speak, release; your speech is transcribed locally and typed/pasted into the active window.

Built with Go, using whisper.cpp's C library with Metal GPU acceleration on Apple Silicon.

Prerequisites

The following is my environment, you can probably build this on Linux, and maybe on Windows without significant changes.

• macOS on Apple Silicon (M-series)
• Go 1.21+
• CMake
• Xcode Command Line Tools (xcode-select --install)

Setup

Clone the repo with submodules:

git clone --recurse-submodules git@github.com:lhk/push2whisper.git
cd push2whisper

Then run the following scripts in order:

1. Download a model

bash go-whisper/download_models.sh

This downloads the default set model (turbo large v3 q5 quantized. That works best for me :) ). You can also pass specific model names, e.g. bash go-whisper/download_models.sh ggml-base.en.bin for a smaller model.

2. Build whisper.cpp

bash go-whisper/build_whisper.sh

Builds the whisper.cpp static libraries with Metal and BLAS acceleration.

3. Build the Go client

bash go-whisper/build_wrapper.sh

Produces the go-whisper/whisper-client executable.

4. Run

cd go-whisper
./whisper-client

Hotkeys

Hotkey	Action
Ctrl + Shift + S	Start/stop recording
Ctrl + Shift + Q	Re-transcribe last recording
Ctrl + Shift + 2	Retype last transcription

Configuration

Set WHISPER_MODEL_PATH to use a different model:

WHISPER_MODEL_PATH=../whisper.cpp/models/ggml-base.en.bin ./whisper-client

CoreML (experimental)

To try CoreML acceleration (runs the transcription on Apple's Neural Engine):

uv venv
source .venv/bin/activate
uv pip install -r whisper.cpp/models/requirements-coreml.txt
bash go-whisper/build_whisper_coreml.sh
bash go-whisper/convert_coreml_model.sh large-v3-turbo
bash go-whisper/build_wrapper_coreml.sh

See the scripts for details on model naming and symlinks for quantized models.