whisperclip

Local, offline voice-to-clipboard dictation for Sway / Wayland. Press a shortcut, speak, press again — your speech is transcribed on-device with whisper.cpp and dropped straight into the clipboard, ready to paste anywhere (an editor, a terminal, an AI coding CLI).

No cloud, no API keys, no account. Your voice never leaves the machine.

Built to dictate long, detailed prompts and code-review feedback instead of typing them word by word.

---

Features

• Toggle dictation — one keybinding starts recording, the same one stops it, transcribes, and copies the text to the clipboard.
• 100% local — whisper.cpp on CPU, or on the GPU via Vulkan.
• Automatic GPU/CPU selection — uses the discrete GPU when on AC power (fast), falls back to the CPU on battery so the dGPU stays asleep.
• Bluetooth-safe — always records from the internal mic, so Bluetooth headphones stay in high-quality A2DP and your music is never interrupted (grabbing the BT mic would force the HFP "call" profile).
• Automatic silence trimming — long thinking pauses are removed with ffmpeg before transcription, so pauses never pollute the text and long clips are faster.
• Text-only history — every dictation is saved as a small .txt (timestamped); the audio is discarded, so nothing bloats your disk.
• Searchable history — a floating fzf window (with a live preview of each transcript) lets you recover and re-copy any past dictation.
• Clean notifications — a single notification card updates in place (recording → transcribing → copied) and dismisses itself.

How it works

              ┌──────────── $mod+Shift+d (toggle) ────────────┐
              ▼                                                 │
  pw-record (internal mic, 16 kHz mono)                         │
              │  stop ─────────────────────────────────────────┘
              ▼
  ffmpeg silenceremove  ──►  whisper.cpp  ──►  wl-copy (clipboard)
                              (GPU on AC / CPU on battery)         │
                                                                  ▼
                                    ~/.local/share/voz-dictado/historial/  (text history)

Requirements

• A wlroots compositor (developed on Sway) + wl-clipboard (wl-copy).
• PipeWire (pw-record, pw-cli).
• whisper.cpp built locally, plus a ggml model (defaults to small).
• ffmpeg (silence trimming), fzf (history picker), a notification daemon such as mako.
• Optional GPU: whisper.cpp built with -DGGML_VULKAN=ON.

On Arch:

sudo pacman -S wl-clipboard pipewire ffmpeg fzf mako
# GPU (optional): vulkan-headers shaderc spirv-headers  + build whisper.cpp with -DGGML_VULKAN=ON

Install

git clone https://github.com/anycodef/whisperclip
cd whisperclip
./install.sh          # symlinks bin/* into ~/.local/bin

Then build whisper.cpp and fetch a model:

git clone https://github.com/ggerganov/whisper.cpp ~/.local/src/whisper.cpp
cmake -S ~/.local/src/whisper.cpp -B ~/.local/src/whisper.cpp/build -DCMAKE_BUILD_TYPE=Release
cmake --build ~/.local/src/whisper.cpp/build -j
sh ~/.local/src/whisper.cpp/models/download-ggml-model.sh small

Finally, add the Sway snippet (see sway/config.snippet) to ~/.config/sway/config and reload.

Usage

Shortcut	Action
$mod+Shift+d	Start recording · press again to stop, transcribe and copy
$mod+Ctrl+h	Floating history picker (search + preview → copy)

From a terminal:

voz-historial              # list past dictations
voz-historial pick         # fuzzy picker (also bound to $mod+Shift+h)
voz-historial <timestamp>  # copy a specific transcript to the clipboard

Configuration

A few constants live at the top of bin/voz-dictado:

• MODEL — path to the ggml model (swap small for base/medium to trade speed for accuracy).
• DGPU_CTRL — the discrete GPU's sysfs power-control path. Adjust the PCI address (0000:01:00.0) to your machine (lspci | grep -i vga).
• Language is set to Spanish (-l es); change it in the whisper-cli call.

License

MIT — see LICENSE.