open_wake_word #

A Flutter FFI plugin for openWakeWord, providing efficient on-device wake word detection using ONNX Runtime.

Features #

• On-device Detection: All processing happens locally on the device for privacy and speed.
• High Accuracy: Uses the power of openWakeWord's models.
• Cross-Platform: Supports Android and iOS (with support for more platforms in progress).
• Fast Execution: Implemented via Dart FFI for near-native performance.

Getting Started #

Prerequisites #

You will need the following ONNX models from the openWakeWord project:

1. melspectrogram.onnx
2. embedding_model.onnx
3. A wake word model (e.g., hey_jarvis.onnx)

Installation #

Add open_wake_word to your pubspec.yaml:

dependencies:
  open_wake_word: ^0.1.0

Asset Setup #

Include the ONNX models in your pubspec.yaml assets:

flutter:
  assets:
    - assets/models/melspectrogram.onnx
    - assets/models/embedding_model.onnx
    - assets/models/hey_jarvis.onnx

Usage #

Initialization #

Initialize the engine by providing the asset paths to your models:

import 'package:open_wake_word/open_wake_word.dart';

bool success = await OpenWakeWord.init(
  melModelAssetPath: 'assets/models/melspectrogram.onnx',
  embModelAssetPath: 'assets/models/embedding_model.onnx',
  wwModelAssetPath: 'assets/models/hey_jarvis.onnx',
);

Processing Audio #

Feed 16kHz PCM audio data (int16) to the engine. For optimal results, use chunks of 1280 samples (80ms).

OpenWakeWord.processAudio(audioData);

if (OpenWakeWord.isActivated()) {
  print("Wake word detected!");
}

double probability = OpenWakeWord.getProbability();
print("Current probability: $probability");

Cleanup #

Don't forget to destroy the engine when it's no longer needed:

OpenWakeWord.destroy();

License #

This project is licensed under the Apache License 2.0 - see the LICENSE file for details. It uses models and concepts from the original openWakeWord project by Kevin Ahrens.