What problem does it solve? Voice interfaces fail at the seams: agents talk over users, hear their own speech, fake a "Listening" status, and collapse every failure into one generic error. This Skill encodes the turn-taking mechanics, noise hygiene, consent lifecycle, and error-repair states needed to build voice UI that survives real microphones, accents, and interruptions. ## Core Features & Use Cases - Barge-in and duck-and-resume engine: duck agent audio on voice onset, decide yield vs resume from the transcript, and resume from the exact paused word after false alarms. - Live interim ink and push-to-talk: render interim transcripts while the user speaks, suppress silence auto-stop while a key is held, and keep open-mic and push-to-talk as parallel modes. - Noise hygiene, echo guards, and truthful status: sustained-onset VAD thresholds, overlap and blip filters for echo, a consent lifecycle, and a state machine that never claims "Listening" without a live mic. - Use Case: When adding voice input to a web app, use this Skill to wire barge-in, A/B speech recognizers on the real user's accent, design the mic-denied recovery card, and ship four distinct error-repair states instead of one "Sorry, I didn't get that." ## Quick Start Use the design-voice-interactions skill to add barge-in, live interim transcripts, and truthful mic status to my voice assistant prototype.
