SeedRealtime

SeedRealtime is a native audio-visual full-duplex Large Language Model (LLM) developed by ByteDance's Seed Team. It represents a significant step towards omni-modal natural interaction by employing a unified architecture that natively fuses audio, video, and text. This enables real-time interaction over continuous multimodal streams, delivering a novel "watch, listen, and speak" experience. The model achieves three core breakthroughs: joint audio-visual understanding, proactive interaction, and natural conversational timing. Unlike traditional cascaded systems that chain separate modules (ASR, VLM, TTS) or many end-to-end models that rely on external Voice Activity Detection (VAD) for turn-taking, SeedRealtime unifies sound, vision, timing, and expression within a single end-to-end model. This allows perception, understanding, decision-making, and expression to run in parallel over continuous audio-visual streams, ensuring that both heard and seen information jointly inform every real-time judgment. It is designed to overcome challenges like conversational rhythm in continuous video and deep contextual understanding, such as disambiguating homophones using visual cues. SeedRealtime has been fully rolled out, pioneering large-scale deployment of audio-visual full-duplex technology in the industry.