Phonotactic structure of the wake word sets the false-wake rate
Aliases: phonotactics · confusable set · false accept · hotword length
What it is
Whether a wake word can survive everyday speech depends first on its phonotactic structure: phoneme count, syllable length, stress pattern, and acoustic distance from high-frequency words, names, and brands. Short, common phrases that overlap ordinary talk raise the keyword detector’s false-accept rate; longer, rarer sequences do the opposite. The issue here is how easily the phrase is collided with as an acoustic template, not how the product handles a false trigger after it fires.
Why it happens
Keyword spotting matches short-time features to a template or a small network on a low-power chip. The decision boundary is held open by positive examples of the target phrase and a mass of negatives (other speech and noise). Two or three common syllables sit in a dense neighborhood: colloquial fragments, television dialogue, and song lyrics all cross the threshold. Adding phonemes, avoiding high-frequency syllable combinations in that language, and staying off slogans that advertisements repeat moves the target out of the dense region. A custom wake word hands the choice to the user and replaces a lab vocabulary with an uncontrolled personal one: a monosyllabic nickname almost necessarily overlaps daily speech. Model capacity is capped by the always-on power budget; a deeper network cannot rescue a template that is itself too short.
Studying it
Estimate false accepts per hour on large negative audio, stratified by phoneme length, syllable count, and minimum edit distance to the lexicon. Confusable-set analysis lists the real phrases that fire the detector. Speaking rate, accent, and channel (handset versus far-field array) belong among the independent variables. Measuring only recall on the official phrase selects words that are memorable and everywhere. Retest after advertising campaigns, because media injects the target phrase itself into the negative distribution.
Where it stops holding
Phonemic uniqueness cannot stop someone from saying the phrase on purpose, or a commercial from playing the slogan verbatim. In multilingual homes, a sequence that is rare in one language may be ordinary in another. Very short wake words are harder to separate from false alarms in noise, but length is not uniformly safer: extra syllables raise misses, and older speakers and children may not produce the full string in one go. Attributing every false trigger to threshold, while ignoring collision density of the word itself, treats tuning as the only lever.
Applying it
- When choosing a phrase, avoid high-frequency syllables in that market and slogans currently on air; decide with negative hours, not with how clever the word sounds.
- If custom phrases are allowed, reject candidates that are too short or homophonous with system commands, and prompt for distinctive combinations of three or more syllables.
- After launch, watch false accepts by region and content source; if a television programme keeps colliding, change the word or add context constraints before simply tightening the threshold.