Glossary
Wake word
A short phrase, such as an assistant’s name, that a device listens for continuously to know when to start processing speech. Wake-word models are small, always-on classifiers trained on many recordings of the phrase.
Training needs the phrase spoken by many people across accents, distances, and noise conditions, plus large amounts of negative audio that teaches the model what to ignore. The two error rates — false accepts and false rejects — are tuned against each other. Because the models run on-device under tight compute budgets, data quality carries unusual weight.
Why it matters
Wake-word work is a recurring brief for collected speech: short utterances, many speakers, wide conditions, clear consent.