Audio as Data
Audio Classification
8 min+75 XPStep 1/10
The ears of Siri and Alexa
Every voice assistant starts the same way: an audio classifier running 24/7, asking one question — did someone say the wake word?
Mel spectrogram in, probabilities over labels out. 'Hey Siri', 'Alexa', 'OK Google' — each is a tiny model living on your device, running on milliwatts, deciding whether to wake the expensive speech-recognition pipeline behind it. That's why wake words are short, and why you can't set a full sentence as one.
insight
Always-listening has a hard energy budget, so the wake-word model is deliberately tiny — which is also why 'a plaything' occasionally wakes Alexa.