ByteDance Seed Releases Native Audio-Video Full-Duplex Model SeedRealtime
ByteDance Seed Team released SeedRealtime, a native audio-video full-duplex large model, now available on the Doubao App. The model unifies audio, video, and text within an end-to-end architecture, processing perception, understanding, decision-making, and expression in parallel, and internalizes turn-taking within the model, replacing external Voice Activity Detectors. Internal human evaluation at ByteDance shows that its rhythm issues are halved compared to cascaded architectures, but no technical report, parameter scale, open-source weights, or API have been released.
