WebSocket-based API for bidirectional voice streaming between the Sinch platform and customer endpoints.
This service enables real-time audio streaming for outbound calls, supporting use cases such as AI-powered voice assistants, interactive voice applications, and custom telephony integrations.
Key Features:
- Bidirectional audio streaming using PCM-encoded audio frames
- Flexible audio chunk sizes (e.g., 20ms frames or larger buffered chunks)
- Stream control commands for audio buffer management and connection keep-alive
- Support for call acceptance, rejection, and busy signaling
Audio Buffering: The platform buffers incoming audio and plays it out to the caller. Use the clear command to discard buffered audio that has not yet been played, enabling real-time interruptions for interactive applications like chatbots. The maximum buffer is 5MBs per call. If exceeded, older data will be discarded.
Connection Flow:
- Platform initiates connection with
ConnectRequest - Client responds with
answer,busy, orreject - If answered, bidirectional audio streaming begins
- Either party can send audio frames as binary WebSocket messages
- Client can send control commands (
clear,heartbeat) during the stream
Connection Loss: There is no reconnection mechanism. If the WebSocket connection is lost for any reason, the call is immediately considered hung up and cannot be recovered.
Latency: End-to-end audio latency is the sum of network round-trip time and platform buffering delay. Sending smaller frames (e.g. 20ms) reduces buffering delay but increases per-frame overhead; larger frames reduce overhead at the cost of higher latency. Audio sent to the platform is queued in the playback buffer before being played out to the caller, so a large backlog of buffered audio will delay playback. Use the clear command to flush unplayed audio when low latency is required, for example to interrupt ongoing playback in conversational applications.
You can download the full AsyncAPI specification here:
👉 Download stream.yml
