wakekit

Train a wake word for any language with TTS, run it on-device in Flutter. No cloud, no network call — the same two frozen ONNX models as the wakekit npm package, plus one ~0.4 MB trained head per word.

Live demo (browser): wakekit.vercel.app

pub.dev wakekit versions are independent of npm wakekit versions. The packages share models and behaviour, not a version number.

A detection means "the wake word was spoken" — nothing more. Whether your assistant should respond, and to what, is an app-level decision.

Audio never leaves the device. Models are bundled in the package; inference is local. There is no network call anywhere in this library.

Install

dependencies:
  wakekit: ^0.1.0

Use

import 'package:wakekit/wakekit.dart';

final models = await loadManifest();
final kit = await WakeKit.load(WakeKitOptions(
  model: models.firstWhere((m) => m.id == 'lada'),
  onHit: (score) => print('wake word heard! $score'),
));

final mic = await listenMic(kit); // mic → detector, all local
// later: await mic.stop(); await kit.dispose();

No mic helper needed? Feed audio from any source: kit.push(float32Samples, sampleRate) — resampling to 16 kHz is handled. kit.configure(threshold: …) retunes live, no reload.

listenMic returns a MicSession. WakeKit.dispose() does not stop the mic — the session owns it. Forgetting stop() leaves the microphone open.

A second listenMic on a kit that is already listening throws StateError.

Example app

flutter/example/ is a manifest-driven picker + start/stop control you can run as-is: cd example && flutter run.

Platforms

	iOS	Android	Windows	Linux	macOS	Web
Inference	yes	yes	yes	yes	no	no
Mic (listenMic)	yes	yes	build-verified, untested	build-verified, untested	—	—

Flutter Web is out of scope (the npm package is the wasm runtime). macOS is not a supported target — the native WakeKit menu-bar app covers macOS; the repo's example keeps a macOS runner only as the local test harness. On Windows and Linux, push() is the supported audio path until a mic smoke test exists.

Consumer floors, imposed by the native ONNX Runtime / mic plugins:

• iOS ≥ 16.0, NSMicrophoneUsageDescription, and use_frameworks! :linkage => :static in the Podfile
• Android minSdk 23, RECORD_AUDIO, and ProGuard -keep class ai.onnxruntime.** { *; }
• Windows / Linux first build downloads ONNX Runtime via CMake FetchContent (offline CI must pre-seed or set ONNXRUNTIME_VERSION)
• Linux mic additionally needs system pulseaudio-utils and ffmpeg

Hot restart (dev): Dart state dies; native ONNX sessions and an open recorder survive until a full restart. The mic indicator can stay lit. Full restart clears it. No runtime mitigation in 0.1.0.

Wake words

The bundled models/manifest.json is the single source of truth — ids, thresholds, labels. Never hardcode a threshold; it belongs to its model.

WakeKit.load on a pending: true entry throws ArgumentError (there is no .onnx to load). Pending entries still appear in loadManifest() so a picker can render them disabled.

Resampling

resampleLinear is a port of the npm helper. Each push resamples its chunk independently — no inter-chunk phase carry — so a non-integral ratio has a small discontinuity per chunk. Fine for these speech features; not for playback.

License

Apache-2.0. The two frozen ONNX models come from openWakeWord (Apache-2.0); see NOTICE. Trained heads were built with synthetic voices (TTS).