Reducing false wakes and reducing missed wakes are opposing goals
Aliases: false reject · miss rate · detection threshold · DET
What it is
Keyword spotting fails in two directions: a false wake (false accept) and a missed wake (false reject). Tighten the threshold and false wakes fall while genuine calls are more often dropped; loosen it and the reverse happens. They are opposing goals on one operating point, not two numbers that further optimisation can both drive to zero. A product has to pick a point for a scene, not promise both rates vanish.
Why it happens
The detector emits a score; a threshold slices score into wake / don’t. A ROC or DET curve describes how the two errors trade. Far-field, noise, and accent worsen both curves at once, but the opposition remains: you still slide on the curve, you do not leave it. Perception is asymmetric—one midnight false wake is more memorable than a miss in the afternoon—so consumer speakers often sit tight; accessibility and industrial settings where a miss is costly sit looser. Personalized hotwords and enrolled speakers can move the whole curve and still leave a trade, only in a better place. A more distinctive phoneme string also moves the curve; it does not cancel the trade.
Studying it
Draw a DET in the target acoustics: false accepts per hour at a fixed miss rate, and misses at a fixed false-accept rate. Report operating points by scene (bedroom at night, kitchen, in-car). Independent variables include threshold, speaker-dependent spotting, and distance. A single “wake success rate” hides the other side. Long-run field false-wake counts must be read with lab misses, or the chosen point will be perfect in the lab and unusable in the living room.
Where it stops holding
Push-to-talk has no such acoustic trade. Cascaded detection (ultra-low-power coarse pass plus a second confirmer) can push one side down at the cost of latency and misses on the second stage. Driving false wakes toward zero without measuring misses yields a quiet device that often will not answer. Children’s voices and post-surgical voices sit on a worse curve; the same operating point is less fair to them.
Applying it
- Publish the operating point by scene: tighter in bedrooms, looser for in-car hands-free, and let people slide within a few notches.
- Any threshold change must regress both false-wake hours and misses for the target population; reporting only one side improving is not allowed.
- Jointly accept with a night-time living room and far-field calls; either side moving past budget is a fail.