OpenWakeWord

This pack registers a wake-word provider for the AgentOS voice pipeline. It runs a local OpenWakeWord ONNX model through onnxruntime-node using audio energy and zero-crossing features.

Installation

npm install @framers/agentos-ext-openwakeword

Peer dependencies: @framers/agentos (>=0.10.40) and onnxruntime-node (^1.17.0).

Usage

import { AgentOS } from '@framers/agentos';

const agentos = await AgentOS.create({
  extensionManifest: {
    packs: [
      {
        package: '@framers/agentos-ext-openwakeword',
        options: {
          modelPath: process.env.OPENWAKEWORD_MODEL_PATH,
          threshold: 0.5,
          keyword: 'hey mycroft',
        },
      },
    ],
  },
});

Configuration

The pack reads no secrets. An explicit modelPath takes precedence over its environment fallback.

| Option | Secret id | Environment variable | Required | Default | | --- | --- | --- | --- | --- | | modelPath | None | OPENWAKEWORD_MODEL_PATH | No | ~/.agentos/models/openwakeword/hey_mycroft.onnx, under the user's home directory | | threshold | None | None | No | 0.5 | | keyword | None | None | No | hey mycroft |

Notes

A local ONNX model file and the runtime peer dependency are needed for detection. The session loads on the first detect call; the pack does not download a model.

detect accepts signed PCM frames as Int16Array. The sample-rate argument is informational and is not validated or used to resample audio. The model receives a float32 tensor named input containing normalized RMS energy and zero-crossing rate. Its input must accept that feature layout. The provider reads the leading probability from the leading output tensor.

A probability strictly greater than threshold returns keyword, confidence and providerId. Other results return null. keyword labels the detection and does not select the model. reset has no effect. dispose releases the loaded session when a release method exists and clears its reference.

License

Apache-2.0