ByteDance's SeedRealtime LLM Joins Conversations Live

ByteDance's Seed team has built a model that keeps a voice attached to a face through an entire noisy dinner party, then jumps into a conversation unprompted when the moment calls for it. That's SeedRealtime, a native audio-visual full-duplex LLM that fuses audio, video and text into one architecture instead of chaining together separate speech-recognition, vision and text-to-speech modules. Seed is pitching it as a break from the cascade approach that most real-time voice-and-camera systems still rely on, where information leaks and latency builds up at every handoff between components.

The bigger claim is about timing. SeedRealtime handles turn-taking inside the model itself, removing the external voice-activity detector that most production stacks depend on to decide when to listen and when to speak. Seed says this adds up to three things: joint understanding across audio and video, proactive interaction without a user prompt, and conversational timing that behaves like a real exchange rather than a queue of turns. The model already runs inside Doubao, ByteDance's consumer assistant, though outside teams have no technical report, no parameter count and no way to plug it in yet.