Wake Word Detection
Listening for one phrase, always, on a budget of milliwatts — where the privacy guarantee is an engineering constraint rather than a promise.
When not to use it
- With a phonetically common phrase. "Alexa" was chosen because the hard 'x' doesn't collide with casual speech.
- Resampling to fix the imbalance. It's 1 in 100,000, continuously. Move the threshold — that's the product decision.
- Benchmarking on clean speech. The negative class is 86,400 seconds a day of real household audio.
- With a cloud confirmation stage, claiming nothing leaves. A false accept means audio left. That's the mechanism behind the accidental-recording stories.
Reach for something else instead
- Push-to-talk — a button. No always-on listening, no privacy question, and people hate it.
- Personalised wake words — speaker verification too. Fixes the TV problem, adds a biometric.
- Open-vocabulary keyword spotting — any phrase, phonetic matching.
- Not having a voice interface — genuinely an option, and it dissolves the whole category.
Sources & further reading
- Chen, Parada & Heigold (2014), Small-footprint keyword spotting using deep neural networks — the model that made it practical.
- Sainath & Parada (2015), Convolutional Neural Networks for Small-footprint Keyword Spotting — the CNN version; the size constraint made concrete.
- Warden (2018), Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition — the open benchmark, and it's honest about the false accept problem.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Measuring accuracy instead of false accepts per hour. The device listens 24 hours a day; accuracy is meaningless here.
- Skipping the cascade. Stage 1 is allowed to be wrong because stage 2 is cheap to run rarely.
- Reading the on-device privacy claim as marketing. The always-on chip has no network path — that's architecture, not policy.
- Assuming the model knows who's talking. It doesn't. That's why a TV advert wakes your speaker.
At a glance
FieldSpeech & Audio
The constraintmilliwatts, tens of KB, always on
The metricfalse accepts per hour, against 86,400 seconds a day of audio
The designa cascade; a tiny greedy model wakes a bigger one
The privacy propertythe always-on chip cannot transmit. Architecture, not policy
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Wake word privacy vs. every other privacy claim here — federated learning's gradients leak, DP's ε is often meaningless, model cards are self-reported. This one is a chip that can't transmit.