Trigger Word Detection
Definition
A sequence modeling task of detecting the exact timing of a specific keyword or phrase (trigger word) in an audio stream, used in always-on voice assistants.
Intuition
The model listens to audio continuously; it needs to detect not just whether a trigger word was said, but when, so the system can activate at the right moment.
Formal Description
Input: spectrogram features from a fixed-size audio window.
Output: binary sequence where for steps immediately after the trigger word ends; label “1” for several consecutive steps after the trigger word (to handle class imbalance and give the model a larger target window).
Architecture: 1D convolutions (local audio feature extraction) → RNN/LSTM layers → sigmoid output per time step.
Training data: synthesis — record trigger word + random background noise + random other audio, mix at random volume levels; class imbalance: far more 0-labels than 1-labels → weight positive examples more heavily.
Applications
“Hey Siri”, “OK Google”, “Alexa” — on-device, low-power, always-on detection; keyword spotting in smart speakers.
Trade-offs
Model must be small enough to run continuously on device (low FLOPs, low memory); latency requirements (respond within ~200ms of trigger); false positive rate must be very low (annoying to users); false negative rate also matters (missed wakes frustrate users).