Silero VAD

A voice activity detection (VAD) plugin for OpenVoiceOS that uses the Silero VAD ONNX model. VAD tells the listener when a chunk of audio contains speech, so the assistant knows when the user stops talking.

The plugin registers two components:

• ovos-vad-plugin-silero, a VADEngine that OVOS uses to detect silence during listening.
• ovos-ww-verifier-silero, a wake word verifier that checks whether the audio right after a wake word detection contains enough speech to confirm the activation.

Both use the same bundled Silero ONNX model, or a custom model path you provide. Input audio must be 16 kHz, 16-bit mono PCM.

Install

pip install ovos-vad-plugin-silero

Usage

Set ovos-vad-plugin-silero as the VAD module in your listener config:

{
    "listener": {
        "VAD": {
            "module": "ovos-vad-plugin-silero",
            "ovos-vad-plugin-silero": {
                "model": "/optional/path/to/model.onnx",
                "threshold": 0.2
            }
        }
    }
}

• model is optional. If you leave it out, the plugin uses the Silero ONNX model bundled with the package.
• threshold is the minimum speech probability (0 to 1) for a chunk to count as speech. The default is 0.2.

To use the wake word verifier instead, set ovos-ww-verifier-silero as a wake word verifier plugin. It takes the same model and threshold options. The verifier default threshold is 0.1.

Related projects

• ovos-dinkum-listener, the OVOS listener that loads VAD plugins like this one.
• ovos-plugin-manager, defines the VADEngine and HotWordVerifier plugin templates this package implements.
• silero-vad, the upstream VAD model this plugin runs.

License

Apache-2.0