Realtime Voice Pipeline
Realtime voice infrastructure

Realtime Voice Pipeline

A full-duplex voice-agent pipeline built on Cloudflare Workers. Speech becomes context, decisions, and streaming audio in one continuous session.

Try the visual demo View architecture
Prototype ready

Voice session

Visual prototype
01
VAD Speech boundary
--
02
STT Live transcript
--
03
LLM Turn decision
--
04
TTS Audio response
--
Press start to preview a realtime turn
Transcript EN
You

Waiting for speech...

Assistant

I found three decisions: ship the beta, keep realtime transcription, and review latency on Friday.

TransportWebSocket
InputPCM16 ยท 16 kHz
Turn latency--
InterruptionReady
One session, four stages

Every turn stays in motion.

The pipeline listens continuously, stabilizes speech, decides when to answer, and streams audio back without exposing internal services.

One duplex session

Audio input, transcript events, turn decisions, and response audio share one WebSocket.

Turn-aware by design

Local voice activity detection coordinates streaming STT, interruption, and response timing.

Cloudflare-native

A Durable Object owns each bounded session while internal services stay behind bindings.

Architecture

Composable models. One realtime contract.

The public protocol coordinates specialized VAD, transcription, language, and speech services while preserving a single client session.

Explore the session
ClientPCM16 audio
WebSocket
Session Durable ObjectVADSTTLLMTTS