better-live-audio

⌘⇧A, speak, ⌘⇧A.

Local dictation for macOS. Granite Speech 4.0 1B via llama.cpp.

Installation

curl -LsSf https://raw.githubusercontent.com/staghado/better-live-audio/main/install.sh | sh

Then:

1. Grant Hammerspoon Accessibility and Microphone access (System Settings, Privacy & Security).
2. Click Hammerspoon menu bar icon, Reload Config.

Usage

Press ⌘⇧A to start, press ⌘⇧A again to stop. Transcript pastes into the focused text field (also left on the clipboard so you can re-paste with ⌘V).

First run downloads the model (~3 GB). Subsequent runs hit the warm server.

Configuration

Edit HF_REPO and CHUNK_SEC at the top of src/better-live-audio/run.sh. Restart the server with pkill -f "llama-server.*granite-speech" to apply.

Uninstall

curl -LsSf https://raw.githubusercontent.com/staghado/better-live-audio/main/uninstall.sh | sh

Requirements

• macOS (Apple Silicon)
• Homebrew

The installer adds:

• llama.cpp
• sox
• Hammerspoon
• jq

How it Works

rec records 16 kHz mono WAV in 30-second chunks. A worker transcribes each chunk as soon as the next one starts, in parallel with the ongoing recording. On stop, the last chunk is finalized and any backlog is drained, then chunk transcripts are joined and copied.

Credits

• Granite Speech 4.0 1B (IBM). GGUF conversion.
• llama.cpp. Granite Speech support: @ReinforcedKnowledge, #22101.
• Hammerspoon.
• Inspired by better-live-text.

License

MIT