vox-core 0.2.1

Voice input layer: captures spoken intent, transcribes locally, injects into active workflow or agent pipeline.

Vox

Vox is the voice input layer for the system. It captures speech via push-to-talk, transcribes it locally with faster-whisper, and injects the text into the clipboard (and optionally into the focused window). No cloud calls; no silent failures.

Install

From PyPI (no clone required; package name is vox-core because vox is taken on PyPI):

uvx vox-core

Or install the tool, then run it (the CLI command is still vox):

pip install vox-core
vox

From source (development or latest):

git clone https://github.com/jeffrichley/vox.git && cd vox
uv sync
uv run vox

Or install in editable mode: pip install -e . (use a venv), then vox.

Pre-built binaries (GitHub Releases)

Packaged binaries for Windows, macOS, and Linux are built on each GitHub Release. Download the archive for your platform (e.g. vox-<version>-windows-amd64.zip, vox-<version>-macos-arm64.zip, vox-<version>-linux-x86_64.tar.gz), unpack it, then run the binary:

• Windows: Unzip the archive, then run vox.exe from inside the vox folder (e.g. .\vox\vox.exe --help).
• macOS / Linux: Unzip or untar the archive, then run ./vox/vox from the extracted directory (e.g. ./vox/vox --help).

On first run, the Whisper model is downloaded automatically (see Transcription model). Binaries are built with PyInstaller and are not signed; you may see a security or Gatekeeper prompt on Windows or macOS.

Configuration

• Config file: ~/.vox/vox.toml. Create the directory if needed: mkdir -p ~/.vox.
• Override path: set VOX_CONFIG to the full path of your config file.
• Env overrides: VOX_HOTKEY, VOX_DEVICE_ID, VOX_MODEL_SIZE, VOX_COMPUTE_TYPE, VOX_COMPUTE_DEVICE, VOX_INJECTION_MODE, VOX_TRAY override the same keys from the file.

Copy the example config and edit:

cp vox.toml.example ~/.vox/vox.toml
# Edit hotkey and optionally device_id, model_size, etc.

Commands

• vox or vox run — Start push-to-talk. By default a small window with a Stop button appears; with use_tray = true in config (or VOX_TRAY=1), a system tray icon is shown instead—click the icon and choose Quit to stop. Press and hold your configured hotkey, speak, release; the audio is transcribed and placed on the clipboard (and optionally pasted into the focused window).
• vox devices — List audio input devices (ID, name, host API). Use this to choose device_id in config.
• vox test-mic [--device ID] [--seconds N] — Record for N seconds, play back the recording, then transcribe and print text. Default 2 seconds. Use to verify mic and model before using vox.

Transcription model (faster-whisper)

• First run: The model is downloaded automatically from Hugging Face (size from config, default base). No system FFmpeg required (PyAV is used).
• CPU: Use compute_type = "int8" in config for lower memory and faster inference.
• GPU: Set compute_device = "cuda" and compute_type = "float16" (or int8) in config. Requires CUDA 12 and cuDNN 9.
• Model size: Set model_size in config (e.g. tiny, base, small, medium, large-v3) for speed vs accuracy.

OS permissions

• Microphone: Required for capture. On Windows, allow app access to the microphone. On macOS, grant Microphone access when prompted.
• Accessibility / input injection: Only needed if you use injection_mode = "clipboard_and_paste" (paste into focused window). On Windows, run the app with normal privileges; on macOS, grant Accessibility permission to Terminal (or the app running vox run) so it can simulate paste.

Definition of Visible Done

A human can verify the MVP by:

1. Install: From repo run uv sync (or pip install -e .).
2. Configure: Copy vox.toml.example to ~/.vox/vox.toml; set hotkey (e.g. ctrl+shift+v) and optionally device_id, model_size, compute_type, injection_mode.
3. Run: Execute uv run vox or vox (or vox run). A small “Vox” window with a Stop button appears, or a tray icon if use_tray is enabled.
4. Trigger: Focus any text field (or leave focus anywhere). Press and hold the configured hotkey, speak a short phrase, release the key.
5. Verify: Paste from clipboard (Ctrl+V / Cmd+V) and see the transcribed phrase. If injection_mode = "clipboard_and_paste", the text also appears in the focused field.
6. Stop: Click Stop in the Vox window (or close the window) to exit.
7. Errors: If mic or model is missing, a clear error message appears (no silent failure).

Development

• Quality gate: just test quality (tests, format, lint, types, docstrings, security checks).
• Tests: just test (pytest). Unit tests under tests/unit/, integration under tests/integration/.