wake word

A predefined spoken word or short phrase whose detection activates a voice interface. The detector, a keyword-spotting model, runs continuously on the device, and only after it fires is audio passed to full speech recognitionspeech recognition, n.: Technology that turns spoken words into written text, as when a phone transcribes a voice message or captions appear automatically under a video., often in the cloud. Wake-word models must be small and power-efficient enough to run always-on, often on a dedicated low-power chip, so they are typically compact convolutionalconvolutional neural network, n.: A kind of neural network built for images and other grid-like data. It slides small pattern detectors across the input, finding edges first, then shapes, then whole objects. or recurrent networks over MFCCsMFCC, abbr.: A short list of numbers, usually about 13, that summarizes the shape of a brief slice of sound in a way that mimics human hearing; for decades the standard input to speech recognizers. or log-mel features, sometimes arranged in a cascade in which a cheap first stage triggers a larger on-device verifier and a final check on a server. Design trades false accepts (waking on similar words or on television audio), usually counted per hour of audio, against false rejects, counted as a share of genuine attempts; some systems also check that the speaker is an enrolled user. Datasets such as Google's Speech Commands (2018) support research on small-vocabulary keyword spotting. Accidental activations mean fragments of private conversation can be recorded, a recurring privacy concern, notably after 2019 reports that contractors reviewed voice-assistant recordings.

False wakes rose when a TV advert used a word that rhymes with our wake word.