I built a sub-500ms latency voice agent from scratch
Outperformed Vapi 2× on latency by treating voice as turn-taking, not transcription.

Multi-speaker voice model with natural interruption and barge-in prevention, genuinely different from turn-taking chatbots.
Enterprise meeting platforms, customer service teams, sales call centers
ChatGPT Realtime · Google Gemini Live · Recall.ai
Outperformed Vapi 2× on latency by treating voice as turn-taking, not transcription.
Honest Pipecat port in Go — single binary beats Python for production voice servers.
Replicates Thinking Machines' multimodal demo on a CPU laptop with commodity models.
The repo actually solves the messy plumbing of live voice agents: modular ASR→LLM→TTS adapters plus an optional PersonaPlex speech-to-speech path, per-agent env overrides, and a Playwright-driven Jitsi bot for room joining. It's a useful MVP for anyone prototyping AI co-hosts, though mixing backends is still manual and PersonaPlex demands extra infra, so it's more pragmatic experiment than turnkey product.
Solves the multi-human voice AI problem without wake words using attention pattern detection.
Another AI sales researcher when Clay and Apollo already dominate this space.