How Do We Make Multilingual Wake Words Reliable?

Training a wake-word model that is actually usable is harder than a one-click training script suggests. A model may work in a quiet room, then miss when the speaker, device, or background sound changes.

Reliable recall comes from the whole chain: language-specific data, realistic augmentation, hard negative mining, evaluation, and a delivery gate. Voicute's pipeline currently supports Chinese, English, Japanese, German, French, Spanish, and Hindi.

Two ways a wake word fails

It can fail to wake when the user says it, or wake when nobody said it. The first is a recall problem; the second is a false-trigger problem. Raising sensitivity can improve one while damaging the other. Stable recognition requires moving the decision boundary with better data and training, not just picking a different threshold.

1. Train for the keyword, not a mismatched feature space

Many systems freeze a pretrained speech embedding and train a small classifier on top. If that embedding was mainly learned from English or speaker-verification data, it may not represent non-English keyword acoustics well. Wake words also depend on the order and timing of syllables.

We train the lightweight keyword model from scratch so it can learn the acoustic shape and temporal order of the target phrase directly.

2. Build each language from its own data

Multilingual support is not translating an English recipe. Languages differ in phonemes, rhythm, stress, connected speech, aspiration, and common confusions. Each language therefore gets its own speech pool, positive variations, negative examples, and evaluation material.

3. Make synthetic speech varied, then add real speech

Multi-speaker TTS provides broad coverage. We vary speaker, pitch, speed, volume, noise, reverberation, and device conditions so the model does not learn one perfect announcer voice.

TTS does not cover every real speaker. Accents, fast speech, personal pronunciation, microphones, and rooms are different. Users can add real recordings to improve adaptation to their actual pronunciation. That is why R0 can validate a TTS-based model, while R1 optimization benefits greatly from real recordings.

4. Evaluate recall and false triggers together

• R0 validates basic keyword recall and supports quick online testing.
• R1 scanning runs the model on music, noise, films, and ordinary speech to collect hard negatives.
• R1 retraining combines those negatives and optional real recordings, then evaluates multiple candidate models.

A model that does not meet the recall and false-trigger requirements does not ship. In our internal delivery evaluation, many language models reach 98%–100% recall; this is a controlled benchmark, not a guarantee for every speaker, device, or environment.

Languages and platforms

Chinese, English, Japanese, German, French, Spanish, and Hindi are supported. Models export to ONNX and TFLite and run offline on Web, Android, ESP32, Python, and Home Assistant. Multiple fixed keywords can also share one model.

Why a script is not enough

The individual ideas are public. The difficult part is keeping the data pool, evaluation standard, hard-negative mining, retraining, and delivery gate connected over time. Every failed candidate improves the next model.