Adventists AI · Full-Duplex Voice

Sophon-Realtime v1

Sophon-Realtime v1 is an open full-duplex voice model family built on a 32B model that listens, decides when to take a turn and writes replies, with three builds (Tag, Flash and E2E) that differ in how the reply is spoken. The page reports VoiceBench 82.8 and Big Bench Audio 81.8%, with several results from the team's own runs, and first-audio latency of about 0.25 s for E2E and 0.31 s for Flash. Code and weights are Apache-2.0, and the Tag and Flash builds serve on one 80 GB GPU via vLLM-Omni.