wake-word-detection

Local wake word detection uses openWakeWord with tuned thresholds and trigger_level to perform on-device keyword spotting without cloud processing. It secures audio buffering with limited retention and clears buffered audio after detection to protect privacy.

Wake word detection acts as a trigger for an auth/verification flow rather than direct privileged actions. This spoofing prevention approach ensures detected keywords initiate secure verification steps before executing sensitive commands.

Production always-on listening requires resource-limited inference, model integrity verification via SHA256 checks, cooldowns to prevent DoS, and proper stream lifecycle management. These guardrails prevent unsafe behavior from false triggers and tampering.

Wake word detection applies to real-time streaming audio pipelines and on-device microphones. It manages the audio stream lifecycle properly to ensure continuous keyword spotting while maintaining resource limits and secure audio buffering.

Local keyword spotting limits audio retention and cleans up buffered audio after detection to reduce sensitive data exposure. By processing entirely on-device without cloud transmission, it maintains audio privacy throughout the detection workflow.