Addressee detection for voice agents: device-directed speech detection that runs before STT, so background speech, side conversations, and the agent's own TTS echo never trigger it. No wake word, model-agnostic, drop-in for LiveKit, Pipecat, ElevenLabs, Twilio, and OpenAI. The layer your VAD and turn detection are missing.

Telnyx media-stream voice-agent reference implementation and a reproducible benchmark harness for measuring voice-agent interruption (barge-in) latency

Realtime voice AI assistant — streaming STT → LLM → TTS over WebSocket, with barge-in and live translation. Runs fully local (Whisper + Ollama + Piper) or with cloud providers.

eval and regression harness for elevenlabs voice agents: audio scoring, llm judges, baseline diffs, closed-loop prompt refinement, ship verdicts

Local-first VAD, barge-in, and turn-taking primitives for interruptible voice agents.

llmpipe is a Go library for linear voice/LLM pipelines: processors pass typed frames downstream (or upstream), with async re-entry, barge-in, idle handling, and pluggable STT/LLM/TTS services.

Minimal real-time voice agent on the OpenAI Realtime API — streaming PCM, server-side VAD, function-calling into local RAG, browser + CLI demos with barge-in

Autonomous real-time voice agent: streaming STT to LLM (function calling) to TTS with sub-500ms latency and barge-in. Offline-first, zero API keys.

STT/LLM/TTS pipeline with barge-in interruption, per-stage latency budgets, and end-to-end voice-to-voice timing. Pipeline overhead p99=1ms; realistic config V2V p50=559ms. 22 tests.

Dialogue runtime for a voice agent: policy as a transition table, barge-in that never resumes, per-stage latency budget at p95, line capacity, speakable-reply guardrails. Dependency-free, no vendor calls.

TypeScript starter kit for production voice agents: WebRTC, telephony, realtime model APIs, STT/TTS pipelines, barge-in, handoff, evals, and post-call workflows.

Open scenario pack and scorer for how a voice agent handles barge-in, interruption, and overlap. Works with Vapi, Retell, Bland, Pipecat, self-hosted LiveKit.

Cancellation-first reference runtime for full-duplex voice agents

Estimate spoken duration of Hindi, English, and Hinglish text from syllable counts - a fast, offline proxy for TTS/voice-bot script timing

Realtime voice AI gateway with turn state, interruption handling, provider fallback, degraded state, audit events, runtime evals, Bun, and TypeScript.

Minimal voice AI agent — STT → RAG → LLM streaming → TTS with barge-in. Built to understand how VoxSetu-class systems work.

Evaluation and regression harness for real-time voice agents. Text evals score words; voice agents fail on timing. Ringdown speaks Twilio Media Streams to your unmodified agent, records both directions with ms timestamps, and gates CI on barge-in latency, dead air, tool correctness and groundedness.