vad2mqtt

Bring the entire OpenVoiceOS VAD ecosystem into Home Assistant as IoT sensors.

vad2mqtt is a real-time Voice Activity Detection bridge built on ovos-plugin-manager. It listens to any microphone, loads any OPM VAD plugin, and publishes speech probability, noise level, and a speech-detected binary sensor to MQTT — with full Home Assistant auto-discovery.

Because OPM plugins are interchangeable, every current and future OVOS VAD plugin instantly becomes a Home Assistant sensor with zero code changes. Install ovos-vad-plugin-silero today, swap to ovos-vad-plugin-webrtcvad tomorrow, or drop in a community plugin next week — vad2mqtt adapts automatically.

A complementary occupancy signal from the microphone you already own.

Why this exists — complementary occupancy detection

Most rooms already have a microphone (smart speakers, SBCs with a cheap USB mic, intercom panels, old phones running Home Assistant Companion). A Voice Activity Detector adds a privacy-respecting complementary occupancy signal with zero extra hardware:

• Complementary, not standalone. VAD is one input among many (PIR, mmWave, door sensors, etc.). It feeds into a probabilistic occupancy estimator such as Area Occupancy Detection. Both the binary Speech Detected sensor and the continuous Noise Level sensor are useful inputs.
• No audio ever leaves the device. Only a 0–1 float (speech probability) and a dB value cross the wire. No transcription, no wake-word, no cloud.
• Lightweight. Silero VAD runs in ~1 ms per 30 ms frame on a Raspberry Pi. CPU usage is negligible.
• Fills gaps PIR leaves behind. PIR sensors need motion + heat. VAD catches presence when someone is sitting still at a desk or behind a PIR blind spot. It also misses people who are silent, which is why it must be fused with other sensors rather than used alone.

What it does

1. Real-time audio — captures microphone input in small chunks (default 30 ms).
2. OPM VAD plugin — loads any ovos-plugin-manager VAD engine. Default is ovos-vad-plugin-silero (ONNX, lightweight).
3. Speech probability — publishes a 0.0–1.0 float every chunk.
4. Noise level — publishes RMS dB at a throttled interval.
5. MQTT + Home Assistant — 4 auto-discovered entities under one device.

Quick start (Docker)

# PipeWire / PulseAudio host (most modern desktops)
vim docker-compose.yml   # set MQTT_HOST
docker compose up -d

# Pure ALSA host (e.g. Raspberry Pi)
# Uncomment the `devices:` line in docker-compose.yml and set ALSA_CARD
vim docker-compose.yml
docker compose up -d

Quick start (pip)

pip install vad2mqtt
vad2mqtt

Home Assistant entities (4 under one device)

Configuration

All settings are environment variables:

Switching VAD plugins — the OPM advantage

Because vad2mqtt uses ovos-plugin-manager, you can swap VAD engines with a single environment variable. Every current and future OVOS VAD plugin is supported:

# Silero VAD — best accuracy, ONNX, ~1 ms/frame on Pi (default)
pip install ovos-vad-plugin-silero
VAD_PLUGIN_MODULE=ovos-vad-plugin-silero

# WebRTC VAD — lighter, no ML model
pip install ovos-vad-plugin-webrtcvad
VAD_PLUGIN_MODULE=ovos-vad-plugin-webrtcvad

# Noise-based VAD — threshold-based, minimal CPU
pip install ovos-vad-plugin-noise
VAD_PLUGIN_MODULE=ovos-vad-plugin-noise

# Future community plugins — just install and point VAD_PLUGIN_MODULE at them

This is the power of the OVOS plugin ecosystem: new VAD research (new ONNX models, new algorithms) ships as a plain pip package. vad2mqtt consumes it immediately with zero code changes.

Architecture

Mic ──► sounddevice ──► 30 ms chunks ──► OPM VAD Plugin
                                               │
                    ┌──────────────────────────┘
                    ▼
            MQTT ──► Home Assistant
            │
            ├── vad_probability (continuous)
            ├── noise_level (throttled)
            ├── model_name (startup)
            └── speech_detected (binary, derived)

Related

• ovos-plugin-manager
• ovos-vad-plugin-silero
• shazam2mqtt — music recognition bridge