wake-word-detection_skill

Overview

This skill implements privacy-first wake word detection using openWakeWord and proven always-listening patterns for a JARVIS-style assistant. It focuses on low-latency, low-resource keyword spotting that runs offline and clears audio immediately to preserve privacy. The workflow is test-driven and optimized for CPU, memory, and false positive reduction.

How this skill works

The detector captures audio from a local input stream, buffers a short sliding window, and runs a lightweight model inference when enough audio accumulates. It applies optimizations such as VAD checks, batch or quantized inference, and memory-mapped model loading to minimize CPU and RAM. On detection the buffer is cleared immediately and a callback is invoked; no raw audio is stored or sent externally.

When to use it

• Offline activation phrase detection for a local assistant (e.g., "Hey JARVIS")
• Always-listening systems where privacy and local processing are required
• Resource-constrained devices needing <5% CPU and <100MB memory
• Prototyping or deploying wake-word models with TDD and verification
• Environments where cloud audio streaming is unacceptable for compliance

Best practices

• Start with tests: write failing tests for detection, privacy, and performance
• Keep audio buffer minimal (<= 1.5–2.0 seconds) and clear immediately after use
• Run VAD to skip inference for silence and reduce CPU usage
• Use quantized or optimized ONNX models and memory-map large files
• Confirm detections with a short confirmation window to reduce false positives

Example use cases

• Trigger a local skill or command processor when user says the activation phrase
• Always-listening kiosk or home assistant that must never send audio off-device
• Embedded systems or Raspberry Pi deployments requiring strict resource caps
• Automated tests that validate detection thresholds, CPU, and memory budgets
• Integrating with a higher-level confirmation layer to lower false alarms