openWakeWord vs Voicute: Frozen Embedding vs From-Scratch TCN
openWakeWord and Voicute solve the same problem with opposite architectures. openWakeWord reuses a frozen, pre-trained feature extractor and trains a tiny head per word. Voicute trains a full network from scratch. That single difference explains most of what you'll experience — stable recall, word-picking, and language support.
The architectures
openWakeWord — frozen embedding + shallow head
Each model is a frozen Google speech_embedding (a 96-dimensional speaker-verification feature) with a small classification head. Training only updates the head. The embedding was built to recognize who is speaking — which means it was trained to ignore what is said. Reusing it for wake-word detection inverts the objective.
Voicute — from-scratch TCN
Voicute trains a temporal convolutional network from zero. The filters learn the acoustic shape of the target word directly, with the convolutional inductive bias (locality, weight sharing, translation invariance) providing data efficiency. No borrowed features, no objective mismatch.
Head-to-head
| Dimension | openWakeWord | Voicute |
|---|---|---|
| Backbone | Frozen speaker-verification embedding | From-scratch TCN |
| Training objective | Reused: speaker ID (ignores content) | Direct: detect the word |
| Recall | 25%–90%, word-dependent | Stable 90%+ |
| Word-picking | Yes — per-word hyperparameters | No |
| Languages | English; others ~0.001 scores | EN / DE / FR / JA / ZH |
| Multi-keyword | One model per word | 2–10 words in one model |
| False-trigger control | Threshold tuning | Multi-stage filtering + cross-head gating |
| Model format | ONNX / TFLite | ONNX (~100–170 KB) |
| Setup | Python / CUDA / notebook | Type a word, auto-train |
| Pricing | Free (open source) | One-time payment per model |
When to choose which
- Choose openWakeWord if you only need English, are happy with the official pre-trained words, and have time to tune.
- Choose Voicute if you need a custom word, a non-English word, or several words in one model — and you want stable recall without a tuning weekend.
For a deeper look at why the frozen embedding is the problem, see why openWakeWord recall is unstable.
FAQ
Is openWakeWord better than Voicute?
For English wake words with the official pre-trained models, openWakeWord is free and convenient. For custom, non-English, or multi-keyword models with stable recall, Voicute's from-scratch TCN avoids openWakeWord's word-picking and 25%–90% recall swings.
What is the difference between a frozen embedding and from-scratch training?
A frozen embedding reuses a feature extractor trained for another task and only trains a small head on top. From-scratch training learns every filter for the wake-word task itself, so it captures phonetic content directly and stays stable across words and languages.
Which is cheaper, openWakeWord or Voicute?
openWakeWord is free but you pay in engineering time: per-word tuning, Python/CUDA setup, and unstable results. Voicute charges a one-time fee per model with no per-device license.
Train a from-scratch wake word →