Skip to main content
Your application connects two independently controlled Telnyx calls to one PyAI Live Translation session. Each person sends their own audio and receives only the translation addressed to them. Start with the calling-platform integration guide. Integration status: this recipe requires your own media bridge. A packaged PyAI interpretation adapter and a completed Telnyx telephone acceptance test are not available yet; the existing @pyai/twilio bridge serves Omni voice agents.

​
Start a stream for each person

Create a Call Control application and configure its call webhooks. Answer the customer call, then create a separate agent call. Associate each authorized call_control_id with the same application conversation and its participant. Keep the calls independently terminated on your bridge so their original audio does not bypass translation through a shared mixer. For each active call, send a server-authenticated POST https://api.telnyx.com/v2/calls/{call_control_id}/actions/streaming_start with this body, using the corresponding bridge URL:
The endpoint and fields are defined by the official Call Control SDK. Use /telnyx/agent for the agent’s separate stream. These bridge URLs are endpoints you implement. Only one streaming/fork operation is supported per call.

​
Translate the audio

Validate the WebSocket handshake’s Ed25519 signature and timestamp, then bind the returned call and stream identifiers to your expected participant. Inspect the negotiated media format before accepting audio. Follow Telnyx’s media streaming contract. For the PCMU configuration above, decode inbound μ-law at 8 kHz and use a stateful streaming resampler to produce PyAI’s 16 kHz PCM16. Wrap it in audio.append for that person’s leg. Convert translated 16 kHz PCM back to 8 kHz μ-law and send a base64 media.payload on the target person’s socket. Keep source and playback queues separate. An optimized L16 path can avoid those application conversions. Telnyx documents its streaming L16 as mono 16 kHz signed little-endian PCM; verify the negotiated format when qualifying that path. Selecting it cannot restore detail lost on a narrowband telephone route. See the official wideband media example.

​
Confirm playback and clear safely

Send a uniquely named mark after each tracked playback boundary. Only a current, uncancelled completion may contribute to playback.report; count samples in PyAI’s 16 kHz units. Telnyx also returns queued marks after clear, so invalidate pending marks before clearing. A cleared mark is not proof that speech played. The documented stream has no separate clear acknowledgement. Qualify a serialized flush barrier before reporting playback.cleared; successful socket writes alone are insufficient. Until then, end translation safely on a transfer or uncertain flush instead of reopening gated input. Retain call IDs, source and translated transcripts, playback outcomes and transfer history in your application. Apply the shared guide’s participant replacement and disconnect sequence, and run its telephone acceptance checks before enabling customer traffic.