openWakeWord Alternative — Stable Recall, No Tuning, Non-English
openWakeWord's custom training is unstable by design. If your word fires half the time and false-triggers the other half, here's the alternative: type a word, get a model that holds 90%+ recall — no hyperparameter tuning, no English-only wall.
openWakeWord's three problems
- Unstable recall. The same word can score 25% or 90% depending on the day — see why openWakeWord recall is unstable.
- Word-picking. The frozen embedding separates some words well and others poorly, so you end up tuning hyperparameters per word.
- English-only. Non-English words collapse to ~0.001 prediction scores because the embedding was trained on English.
What to look for in an alternative
- From-scratch training, not a frozen embedding borrowed from a different task.
- No per-word tuning — one workflow that works for every word.
- Real non-English support — German, French, Japanese, Chinese.
- Portable, offline ONNX — not locked into a specific SDK.
- No per-device fees — pay once, deploy anywhere.
Voicute as the alternative
Voicute is an online custom wake word platform built on a different architecture: a from-scratch TCN. You type a word; the system synthesizes training data, trains the model, and returns a standard ONNX file in 10–20 minutes.
| openWakeWord | Voicute | |
|---|---|---|
| Architecture | Frozen speaker-embedding + head | From-scratch TCN |
| Recall | 25%–90%, word-dependent | Stable 90%+ |
| Hyperparameter tuning | Required per word | None |
| Languages | English (others ~0.001 scores) | EN / DE / FR / JA / ZH |
| Model artifact | ONNX / TFLite | ONNX (~100–170 KB) |
| Pricing | Free (open source) | One-time payment per model |
One note on the newer livekit-wakeword: it solves the "one-command training" ergonomics and claims better recall than openWakeWord, but it still synthesizes with espeak-ng (a low-quality, English-centric TTS) and reuses the openWakeWord ONNX pipeline — so non-English quality and word-picking remain open questions. Voicute trains from scratch, which is a different category.
FAQ
What is the best openWakeWord alternative for custom wake words?
Voicute is a from-scratch-training alternative: type a word, get a stable 90%+ recall ONNX model in 10–20 minutes, with native English, German, French, Japanese, and Chinese support and no per-word tuning.
Why is openWakeWord's custom training unstable?
openWakeWord builds each word on a frozen English speaker-verification embedding that was trained to ignore what is said. That objective mismatch causes word-picking and recall that swings from ~25% to ~90%.
Does an openWakeWord alternative support non-English wake words?
Voicute does. Because each model is trained from scratch, German, French, Japanese, and Chinese wake words hold the same 90%+ recall as English.
Is the ONNX model compatible with Home Assistant?
Yes. Voicute outputs a standard ONNX model that runs through the open-source onnx-wakeword engine as a Wyoming service for Home Assistant, or on-device for Android, Linux, and ESP32.
Train your custom wake word →