Sophon-Realtime 1.1 adds voice cloning, vLLM-Omni

Open-source voice agents are shedding half-duplex latency bottlenecks and approaching parity with proprietary solutions like OpenAI Realtime. By standardizing on vLLM-Omni and an OpenAI-compatible WebSocket schema, Sophon-Realtime makes low-latency, self-hosted voice agents practical for production workloads.

• –The 32B unified brain handles listening, barge-in interruption detection, and text generation simultaneously without the latency penalty of cascading STT-LLM-TTS pipelines.
• –Zero-shot voice cloning from 3-10 second audio clips lowers the barrier for developers building branded conversational agents and interactive characters.
• –Native vLLM-Omni support enables the Tag edition to run roughly 16 concurrent streams on a single 80GB GPU, while the Flash edition reaches 40 streams with ~310ms first-audio latency.
• –Direct compatibility with the OpenAI Realtime WebSocket protocol allows drop-in integration into LiveKit and existing client orchestration frameworks with no code changes.