GenPark AI Agent Skill - Voice Turn-Taking & Endpointing Detector

Acoustic voice activity detection (VAD), dynamic silence endpointing, and barge-in interruption arbitrator for real-time conversational voice agents.

Verified by GenPark AI and compatible with Model Context Protocol (MCP).

Architecture Diagram

graph TD
    A[Real-time PCM Audio Stream] --> B[Energy & VAD Frame Analyzer]
    B --> C{Agent Currently Speaking?}
    C -->|Yes & Speech Energy > Threshold| D[Barge-In Interrupt Triggered: Cut Agent Audio]
    C -->|No| E[Accumulate Speech & Silence Frames]
    E --> F{Silence Duration >= Endpointing Threshold?}
    F -->|Yes| G[Decision: Turn Completed -> Dispatch LLM Synthesis]
    F -->|No| H[Decision: Continue Listening]

Related MCP server: voice-mcp-server

Features

• Low-Latency VAD Heuristics: Operates on sub-frame time slices for immediate interruption handling.

• Dynamic Endpointing: Automatically balances conversational fluidity against premature turn cutoff.

• Zero External Dependencies: Pure Python standard library implementation.