protoVoiceFull-duplex voice agent

Whisper STT + vLLM + Fish Audio / Kokoro TTS on Pipecat — sub-200ms TTFA, speak-while-thinking, push-interrupt.

🎤

Streaming Pipeline

Mic → Silero VAD → Whisper Turbo → Qwen → Fish / Kokoro → Speaker. Each stage streams so audio plays while the LLM is still generating.

🔁

The agent speaks filler while tools run, and interrupts the user to deliver long-running tool results. No polite waiting.

🐟

Fish Audio S2-Pro (default, voice-cloning, 44.1 kHz) or Kokoro 82M (low-latency preset voices) — swap via env.

🔌

Dispatches to other protoLabs agents via A2A, starting with Ava as the orchestrator. Optional inbound A2A endpoint so other agents can call us.

This site follows the Diátaxis framework:

Section	Purpose	Start here if you…
Tutorials	Learning-oriented walkthroughs	Are new to protoVoice
How-To Guides	Task-oriented procedures	Need to accomplish something specific
Reference	Technical descriptions	Need exact details on an API, env var, or frame type
Explanation	Understanding-oriented discussion	Want to understand how and why things work