wakekit 🌼

🎙️ Live demo

Train a wake word for any language with TTS, run it in the browser. No GPU, no torch, no cloud — three small ONNX models on a worker thread.

Browser ports of openWakeWord's runtime already exist. What didn't exist is the rest: openWakeWord's sample generator ships one English TTS checkpoint, so training a wake word in Thai — or anything else — was off the table. wakekit is the whole pipeline:

• Runtime — streaming detection in a Web Worker on onnxruntime-web (wasm, single-threaded: works on any static host, no cross-origin isolation). Audio never leaves the tab.
• Training — TTS corpus generation for your language, a numpy-only trainer (~110k params, CPU, minutes), and an eval harness that reports the two numbers that matter: recall and false fires per minute.
• Model zoo — every wake word is one ~0.4 MB head + one manifest entry. Shipped today: ละดา (Thai), trained end-to-end with this repo. Next: อีดี, jarvis, yours.

Try it

Live demo: wakekit.vercel.app — pick a wake word, press start, say it.

Or run locally:

npm install
npm run dev        # → http://localhost:5173 — pick a wake word, press start, say it

The test page shows the live score trace, per-model thresholds, detections, and everything on this page — picker included — is driven by models/manifest.json.

Use the library

npm install wakekit

The package ships the models too (node_modules/wakekit/models/) — copy them into your app's static dir, or skip hosting entirely and point base at the CDN: https://cdn.jsdelivr.net/npm/wakekit/models/. Also serve ort-wasm-simd-threaded.{mjs,wasm} from node_modules/onnxruntime-web/dist/ (see vite.config.ts for the dev-server/build recipe).

import { WakeKit, listenMic, loadManifest } from 'wakekit';

const models = await loadManifest('/models/');

const kit = await WakeKit.load({
  model: models.find((m) => m.id === 'lada')!,
  base: '/models/',                       // where the .onnx files are served
  onHit: (score) => console.log('wake word heard!', score),
});

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

A detection means "the wake word was spoken" — nothing more. Whether your assistant should respond (vs. the word merely occurring mid-conversation) is an app-level decision; gate it with your own logic.

Wake words

| id | label | lang | threshold | trained on | |---|---|---|---|---| | lada | ละดา | th | 0.95 | synthetic Thai voices — held-out: recall 100%, 0 false fires |

Adding one is a .onnx file plus one models/manifest.json entry — the demo and the library pick it up with no code change. The threshold travels with the model, because each head puts its positives and negatives in a different place; ship the number you measured, not a global constant.

Train your own — any language

make help          # the pipeline, documented
make train-lada    # the worked example, end to end

Four steps: TTS corpus (positives + rhyming trap words in your language + everyday speech) → featurize → train (numpy) → measure on held-out voices. Full guide with the ละดา case study: docs/training.md.

Honest caveat: TTS evals prove the chain and unseen-speaker recall; they cannot predict false fires per hour in a real noisy room. Collect real captures, feed them back, retrain — the scripts support that loop.

Beyond the browser

The models are plain ONNX — the same three files run under any ONNX Runtime binding. Worked examples in Python, Node.js, Rust, Go, C#, Java, C++, and Swift: docs/other-languages.md.

How it works

mic (any rate) ──resample──► 16 kHz f32 ──► melspectrogram.onnx   (frozen)
                                        ──► embedding_model.onnx  (frozen, Google speech-embedding)
                                        ──► <word>.onnx           (trained head, ~0.4 MB)
                                        ──► score every 80 ms ──► hit when ≥ threshold

npm run selfcheck drives the real worker over the real models and asserts labelled clips when present (eval/clips/pos_*.wav must fire, neg_*.wav must not).

Credits & license

Apache-2.0 (see LICENSE, NOTICE) — free to use, modify, and redistribute, including commercially.

| component | author / source | license | |---|---|---| | openWakeWord — feature pipeline + frozen melspectrogram.onnx | David Scripka | Apache-2.0 | | embedding_model.onnx — derived from Google speech-embedding | Google | Apache-2.0 | | ONNX Runtime Web — wasm inference | Microsoft | MIT | | highlight.js — demo code highlighting | highlight.js contributors | BSD-3-Clause | | thinking-orbs — demo wake-pill orb | | MIT | | Trained heads (models/lada*.onnx) — synthetic speech (TTS) | wakekit | Apache-2.0 |