Hear the turn
TEN-VAD runs beside the session, finding speech boundaries without another network hop.
Cloudflare-native voice infrastructure
The voice stack, in one continuous session.
Listen, transcribe, reason, and speak without stitching four public APIs together. One realtime contract carries the entire turn.
WebSocket control · PCM and WebRTC media · up to 60-minute sessions
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
TransportWebSocket
InputPCM16 · 16 kHz
Turn latency--
InterruptionReady
TransportFull duplex
Session ownerDurable Object
Speech boundaryLocal VAD
Public shapeOpenAI Realtime
One living turn
Every stage begins before the last one feels finished.
The session keeps audio and meaning in motion. Boundaries stay explicit, while the user experiences one conversation.
Stream the meaning
Realtime transcription starts on speech and emits immutable transcript deltas.
Decide the response
Turn policy chooses whether to wait, acknowledge, answer, or interrupt.
Speak while thinking
Provider-neutral audio decoding streams raw response audio back into the session.
Built as a system
State lives close to the conversation.
A Durable Object owns each session. Local WASM handles signal processing. Provider Workers remain composable and private behind service bindings.
Hibernation preserves the socket while the object sleeps.
Session topologyPrivate service composition
ClientWebSocket / WebRTC
control + media
RealtimeVoiceSessionDurable Object
RNNoiseSpeexDSPTEN-VAD
STTLLMTTS
Realtime event stream
→ input_audio_buffer.append
← input_audio_buffer.speech_started
← conversation.item.input_audio_transcription.delta
← response.output_audio.delta
← response.done Familiar outside. Composable inside.
A public protocol your client already understands.
The gateway exposes an OpenAI-compatible Realtime WebSocket while specialized Workers own transcription, reasoning, and speech generation.
Configure an agentTechnical foundation
Less orchestration in your client. More control at the edge.
One Durable Object
One bounded object owns conversation order, quota state, audio flow, interruption, and transparent hibernation.
DSP at the edge
RNNoise, SpeexDSP, and TEN-VAD run with the Worker. Audio is prepared before it reaches transcription.
Bindings, not public hops
STT, LLM, image analysis, and TTS stay behind explicit Cloudflare service bindings.
OpenAI-compatible events
Use a familiar Realtime WebSocket shape while keeping model routing and provider credentials private.
Voice should feel like a conversation,
not a chain of requests.