ByteDance's Seed team has introduced SeedRealtime, a native audio-visual full-duplex large language model that fuses audio, video, and text in a single unified architecture. Unlike traditional cascaded stacks of ASR, VLM, and TTS modules, SeedRealtime runs perception, understanding, decision-making, and expression in parallel to enable real-time interaction over continuous multimodal streams.
- The model replaces external voice-activity detectors with internal turn-taking mechanisms.
- It demonstrates identity binding across modalities, proactive speech from held instructions, and correction based on visual state.
- Human evaluation reports indicate that pacing issues were halved compared to cascaded stacks.
- SeedRealtime is currently live inside ByteDance's Doubao app but lacks a technical report, open weights, or public API endpoints.
The release serves as a validated reference architecture for real-time voice-plus-camera products, establishing a new benchmark for omni-modal interaction despite the absence of immediate third-party integration options.