Telnyx is a carrier-grade communications platform that voice agents use to place and receive phone calls, and it also sells the speech and inference layers an agent needs. For a bring-your-own-agent setup, Programmable Voice media streaming forks live call audio to your WebSocket server; for a managed setup, Voice AI Agents run the whole STT, LLM, and TTS loop on Telnyx infrastructure.
Media streaming starts from a Dial or Answer command with a streamurl and a streamtrack of inbound, outbound, or both. The socket receives connected and start events with call metadata and media format, then media events carrying base64 RTP payloads with chunk numbers for reordering, plus dtmf and stop events. Setting streambidirectionalmode: rtp lets your server send base64 RTP or MP3 audio back into the call, with mark and clear messages to manage the playback queue. Codecs include PCMU (8 kHz default), PCMA, G722, OPUS, AMR-WB, and L16, the last one added to reduce latency for voice-agent integrations. One bidirectional stream is allowed per call.
Voice AI Agents can be defined through the API or the no-code AI Assistant Builder in Mission Control Portal using natural-language instructions. STT routes to Deepgram, Whisper, AssemblyAI, Speechmatics, or Soniox through one API; TTS offers 1,300+ voices across 10+ languages from ElevenLabs, Rime, MiniMax, Resemble AI, AWS, Azure, and Telnyx Ultra; and an OpenAI-compatible Inference API serves open-weight models on Telnyx-owned GPUs co-located with its telephony points of presence. Telnyx operates as a licensed carrier in 45+ countries.
Features
- Media streaming: bidirectional WebSocket audio for Programmable Voice calls with start, media, DTMF, mark, clear, and stop events
- Codec choice: PCMU, PCMA, G722, OPUS, AMR-WB, and L16 for RTP streams
- Voice AI Agents: managed agents built from natural-language instructions via API or the AI Assistant Builder
- STT and TTS routing: multiple speech providers behind one API, including bring-your-own ElevenLabs voices
- Inference API: OpenAI-compatible endpoints for models such as DeepSeek, MiniMax, and Kimi on Telnyx GPUs
- Agent tooling: dynamic variables, call transfer, memory, calendar and CRM integrations, and MCP servers
- Carrier network: direct call origination in 45+ countries with STIR/SHAKEN attestation for eligible US outbound calls
- Framework examples: reference integrations for OpenAI speech-to-speech and a Pipecat chatbot
- Adjacent primitives: SIP trunking, SMS, WhatsApp, WebRTC, number lookup, and edge compute under the same account
