Keyword Spotting (KWS) is the first stage of any voice-interaction product: a compact neural model runs locally on the audio SoC, continuously monitoring the microphone feed and deciding whether the preset wake word has been spoken. Unlike cloud ASR, KWS performs a lightweight classification — "wake word or not" — so it responds within milliseconds, works offline, and consumes minimal compute. For TWS earbuds, smart speakers, and call-center headsets, KWS quality directly defines whether the product feels instant or unresponsive, making it the single most user-visible parameter of a voice feature.
Battery-powered devices cannot afford to run a neural network around the clock. The proven architecture is hierarchical: a microamp-level Voice Activity Detector (VAD) listens first and only wakes the KWS engine when speech-like energy appears; once the wake word is confirmed, full ASR and network requests start. On typical Bluetooth audio SoCs this duty-cycled approach keeps the standby increment in the microamp-to-milliamp range, effectively invisible in battery-life figures. In our reference designs we tune the VAD threshold, listening window, and SoC sleep policy together — an aggressive sleep schedule that saves power but misses wake words is worse than no feature at all.
Most production KWS pipelines extract MFCC or Mel-filterbank features and feed them into a compact CNN or CNN+GRU classifier weighing tens to a few hundred kilobytes, small enough to execute on the audio SoC itself or a low-power co-processor. The engineering trade-off is three-way: a smaller model lowers power but degrades noise robustness; a deeper network improves accuracy but raises inference cost and RAM footprint. Custom wake words additionally require training corpora that span genders, accents, and distances. Typical acceptance criteria are above 95% wake-rate in quiet conditions with a strictly bounded false-wake count per 24 hours.
Real-world behavior depends on more than the model. False wakes originate from TV dialogues, song lyrics, and babble noise, and are controlled through negative-sample training combined with dual-microphone beamforming. Far-field wake-up hinges on microphone sensitivity, SNR, and echo cancellation: when the device itself is playing music, acoustic echo cancellation (AEC) must run ahead of the KWS stage, otherwise wake-while-playing simply does not work. In production we add standardized wake-word and noise corpora to the line test plan so every batch is verified against the same wake-rate and false-wake criteria.
Q: How is keyword spotting different from speech recognition (ASR)?
A: KWS only decides whether a preset wake word appeared — a low-compute local classification task. ASR converts full speech into text and needs far more compute or cloud support. Products typically use KWS to wake up first, then launch ASR for the follow-up command.
Q: How can false wake-ups be reduced?
A: Train with extensive negative samples such as TV audio and similar-sounding phrases, use a dual-microphone array to focus on the front direction, and apply a two-stage decision where VAD filters out non-speech segments before the KWS model makes the final call.
Q: How is power controlled for wake-up on battery devices?
A: Use the VAD-plus-KWS hierarchy so the neural network runs only briefly when speech is present, and combine it with the SoC's low-power listening mode and fast sleep entry. The standby increment stays in the microamp-to-milliamp range, which is negligible for battery life.
Shenzhen Liwei Electronic Technology Co., Ltd has over 10 years of experience in audio PCBA design, serving 15+ well-known brands and 300+ customers with complete chip-and-PCBA solutions for voice-enabled headsets, smart speakers, and call-center headphones, including low-power wake-up tuning and production test support. Contact us today for a custom audio solution — Request a Quote.
Shenzhen Liwei Electronic Technology Co., Ltd has over 10 years of experience in audio PCBA solution design, serving 15+ well-known brands and 300+ customers. Contact us for custom voice-enabled audio solutions.
Keywords: keyword spotting, KWS, voice wake-up, VAD, low power audio