Linux-Voice-Assistant

An experimental voice satellite software for Home Assistant remote voice control and interaction.

This project enables you to build a Linux-based voice assistant designed to use Assist for Home Assistant. It allows you to create your own smart speaker that runs on any x64 or ARM64 hardware capable of handling local audio processing (using PulseAudio).

Unlike simpler voice satellites that run on microcontrollers with very limited compute power, this setup can perform local wake word detection (OWW/MWW) and process some data on-device.

Because it runs on a full Linux system and offers access significantly more local computing resources for additional features and other integrations on the same satellite, this approach also provides greater flexibility for customization (such as for example experiment with using PipeWire).

Features

• Works with Home Assistant using the ESPHome protocol/API (via aioesphomeapi)
• Feature local on-device wake word detection using integrated OpenWakeWord or MicroWakeWord
• Supports multiple wake words and languages
• Supports multiple architectures (linux/amd64 and linux/aarch64)
• Automated builds with artifact attestation for security
• Supports announcments, start/continue conversation, and timers
• Tested and works with Python 3.11 and Python 3.13.
• Prebuild docker image available on GitHub Container Registry
• Prebuild Raspberry Pi image
• Supports Websocket API for peripherals (e.g. buttons, LEDs, etc.) to integrate with the voice assistant

Planning for future features is discussed in Discussion #495.

Requirements

• Microphone: Device must support 16kHz mono audio
• CPU: 1Ghz
• Memory: min. 512MB
• Storage: The OS and software is around 4GB
• OS: linux/amd64 or linux/aarch64

A more extensive list for possible compatible hardware can be found in the PiCompose documentation but basically any microphone that works with PipeWire (multimedia framework for Linux) can in theory be used for voice input with the prebuild image from there, you should however preferably use a far-field microphone-array solution if want better result.

Two solutions recommended for setups today is:

• use a Raspberry Pi Zero 2W (Single Board Computer with built-in WiFi) in combination with the Satellite1 Hat Board
• use at least a Raspberry Pi 3 with the Respeaker Lite. The Respeaker Lite currently has a problem with the Zero 2W.

Those mic-boards have microphone-array designed for far-field voice capture with the added benefit of using an onboard XMOS DSP microcontroller with custom firmware which does advanced audio pre-processing for microphone cleanup that result in very good voice recognition capabilities (as it runs algorithms for Noise Suppression, Acoustic Echo Cancellation, Interference Cancellation, and Automatic Gain Control).

Alternatively if on a lower budget then suggest could use other microphone-array boards like for example the reSpeaker 2-Mics Pi HAT V2.0 (which uses a much more basic audio codec chip).

Usage

Installation

Assist Satellite app for Home Assistant OS

For HA OS, we provide a finished Assist Satellite app (formerly add-on), which uses the Linux Voice Assistant runtime to turn your HA host into a voice satellite.

[!NOTE] For now you first have to add the OHF-Voice apps repo manually to the App Store repository inside Home Assistant before you can install it.

Later you will be able to install it directly from the official add-on repository (but it is not yet published publicly there):

Once installed, the satellite is automatically discovered by Home Assistant via the ESPHome integration.

Raspberry Pi prebuilt image

For Raspberry Pi users, we provide a prebuild image that can be flashed to a SD card. See PiCompose.

Docker / bare metal

For all other users, we have different installation methods available (Docker, systemd), each with its own dedicated instructions. See Linux-Voice-Assistant - Installation.

Parameter overview

💡 Note: There is an environment variable for each parameter if you use docker or systemd based setup.

usage: __main__.py [-h] [--name NAME] [--audio-input-device AUDIO_INPUT_DEVICE] [--list-input-devices] [--audio-input-block-size AUDIO_INPUT_BLOCK_SIZE] [--audio-output-device AUDIO_OUTPUT_DEVICE] [--music-output-device MUSIC_OUTPUT_DEVICE] [--list-output-devices] [--wake-word-dir WAKE_WORD_DIR]  [--mic-auto-gain] [--mic-noise-suppression]
                   [--wake-model WAKE_MODEL] [--stop-model STOP_MODEL] [--download-dir DOWNLOAD_DIR] [--refractory-seconds REFRACTORY_SECONDS] [--wakeup-sound WAKEUP_SOUND] [--timer-finished-sound TIMER_FINISHED_SOUND] [--processing-sound PROCESSING_SOUND]
                   [--mute-sound MUTE_SOUND] [--unmute-sound UNMUTE_SOUND] [--preferences-file PREFERENCES_FILE] [--host HOST] [--network-interface NETWORK_INTERFACE] [--mac-address MAC_ADDRESS] [--port PORT] [--enable-thinking-sound] [--listen-during-wake-sound] [--debug]

Parameter	Description	Default
--name	Name of the voice assistant device (required)	Autogenerated (lva-MAC-ADDRESS)
--audio-input-device	Soundcard name for input device	Autodetected
--audio-input-block-size	Audio input block size in samples	1024
--audio-output-device	mpv name for output device, use a pulse/... name from --list-output-devices	pulse
--music-output-device	mpv name for the music/media output device	--audio-output-device
--mic-volume	Control microphone volume	100
--mic-auto-gain	Add WebRTC Gain to Mic	0
--mic-noise-suppression	Add WebRTC Noise Suppression to Mic	0
--audio-input-channels	Number of microphone audio channels to stream	2
--wake-word-dir	Directory with wake word models (.tflite) and configs (.json)	wakewords/
--wake-model	ID of active wake word model	okay_nabu
--stop-model	ID of stop model	stop
--download-dir	Directory to download custom wake word models, etc.	local/
--refractory-seconds	Seconds before wake word can be activated again	2.0
--continue-conversation-delay	Delay before mic opens for continued conversation	0.5
--timer-max-ring-seconds	Seconds after which the timer stops ringing	900.0
--wakeup-sound	Sound file played when wake word is detected	sounds/wake_word_triggered.flac

💡 Note: There are detailed explanations on the controlled entities from device page in the configuration file.

💡 Note: How the assistant behaves in everyday situations (music ducking during conversations, announcements, timers, stop word, mute, volume) is described in the behavior guide.

Build Information

Image builds can be tracked in this repository's Actions tab, and utilize artifact attestation to certify provenance.

The Docker images are built using GitHub Actions, which provides:

• Automated builds for different architectures
• Artifact attestation for build provenance verification
• Regular updates and maintenance

The documentation for the build process can be found in the GitHub Actions Workflows file.

Development

System Requirements

System packages (Linux):

• libmpv-dev (sudo apt install libmpv-dev)
• PulseAudio/PipeWire (sudo apt install pulseaudio pipewire)
• ALSA utils (sudo apt install alsa-utils)

Python:

• Python 3.11+ must be installed on your system (python3 --version should show 3.11+)
• On Ubuntu/Debian: sudo apt install python3.11 python3.11-venv python3.11-dev

VS Code Setup

VS Code development uses a local Python virtual environment (.venv/):

1. Open workspace in VS Code
2. Accept Workspace Trust (bottom-left status bar)
3. Install recommended extensions when prompted:
   • ms-python.python - Python language support
   • ms-python.vscode-pylance - Python language server
   • kilo.kilocode - Kilo AI assistant
4. Terminal → New Terminal (or Ctrl+`) - Opens integrated terminal (uses .venv when available)
5. Run ./script/setup --dev to create .venv and install dev dependencies
6. VS Code automatically detects .venv/bin/python as the Python interpreter

Available VS Code Tasks:

• Ctrl+Shift+B - Run Setup or Linter
• Ctrl+Shift+T - Run Tests
• Ctrl+Shift+P → "Tasks: Run Task" - Show all tasks (Setup, Linter, Tests, Run App)

Code Quality Checks

The project uses the following tools to ensure code quality:

• Black: Code formatting (200 characters per line)
• isort: Import sorting compatible with Black
• flake8: Style and syntax checks
• pylint: Code quality checks
• mypy: Static type analysis

Docstrings are mandatory in Sphinx style with :param: (and :raises: where relevant). Google-style (Args:) and bullet-style docstrings are not accepted:

def my_function(param1: str, param2: int) -> str:
    """
    Brief one-line description of the function.
    :param param1: Description of what param1 is used for.
    :param param2: Description of what param2 is used for.
    """

Setup

To use the development tools (linting, testing, etc.), you need to install the required dependencies:

./script/setup --dev
source .venv/bin/activate

Linting Commands

Run all linting checks

./script/lint...

Individual linting commands (with auto-fix support)

Script	Description	Auto-fix Available?
./script/lint_black	Checks Python code formatting with Black	Yes, use--auto flag
./script/lint_flake8	Runs style and syntax checks with flake8	No
./script/lint_isort	Checks import sorting with isort	Yes, use--auto flag
./script/lint_mypy	Runs static type analysis with mypy	No
./script/lint_pylint	Runs code quality checks with pylint	Yes, use--auto flag

Examples

Run a specific lint check:

./script/lint_black

Auto-fix formatting issues (Black + isort):

./script/lint_black --auto
./script/lint_isort --auto

Testing

Run the test suite:

./script/tests

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.