wakekit library

wakekit — wake word detection on-device, openWakeWord-compatible.

final models = await loadManifest(); final kit = await WakeKit.load(WakeKitOptions( model: models.firstWhere((m) => m.id == 'lada'), onHit: (score) => print('wake! $score'), )); final mic = await listenMic(kit); // later: await mic.stop(); await kit.dispose();

A detection means "the wake word was spoken" — nothing more. Whether your assistant should respond, and to what, is an app-level decision (same contract as the npm package's README). Audio never leaves the device: every model runs locally, no network call exists in this package.

Classes

MicSession

A running microphone capture started by listenMic. The ecosystem idiom for "a thing you stop" (StreamSubscription.cancel, Timer.cancel) — an object handle, not a bare function. WakeKit.dispose() does NOT stop this: the session owns the mic, so forgetting stop leaves the microphone open until it's called (documented limitation, see the spec's edge table).

WakeKit

Loads the three ONNX models for one wake word and streams audio through them. One instance per active wake word; dispose() when done, then create a new one to switch words.

WakeKitOptions

Options for WakeKit.load. Callback-based, mirroring WakeKitOptions in src/index.ts:47-62 — a Flutter-native pattern, and it keeps this README teachable side by side with the npm one.

WakeModel

One entry in models/manifest.json — the model registry the manifest-driven picker (FR-6, FR-12) is built from. Never hand-construct thresholds or ids: read them from loadManifest.

WakeModelEval

Held-out measurement (unseen speakers) behind a model's threshold — see scripts/eval.mjs.

Functions

listenMic(WakeKit kit) → Future<MicSession>

Requests mic permission, opens a 16 kHz mono PCM16 stream, and pushes it into kit. Throws StateError if permission is denied or if kit already has an active mic session — two recorders feeding one pipeline would corrupt its mel/embedding buffers silently, not loudly. The requested rate is a request, not a fact (record reports no negotiated rate) — chunks still route through push(_, 16000), so a mis-rated device is a documented limitation, not a crash.

loadManifest() → Future<List<WakeModel>>

Reads the bundled models/manifest.json — the same file the npm package and demo page ship, synced byte-for-byte by scripts/sync-flutter-assets.mjs. Never hand-maintain a parallel list of ids/thresholds/labels; this is the single source of truth (CLAUDE.md).

resampleLinear(Float32List input, int inRate, int outRate) → Float32List