ByteDance's Seed research team released SeedRealtime on August 5, a native audio-visual full-duplex large language model that fuses sound, vision, and text within a single unified architecture. Unlike cascaded systems that chain separate modules for speech recognition, vision, and text-to-speech, SeedRealtime runs perception, understanding, and response generation in parallel over continuous multimodal streams.
The model has already been deployed in ByteDance's Douyin and Doubao consumer apps, which the company claims is the first large-scale rollout of full-duplex audio-visual technology.
Three capabilities distinguish SeedRealtime from earlier approaches. First, joint audio-visual understanding lets the model resolve ambiguities by cross-referencing what it hears with what it sees -- for example, disambiguating homophones by analyzing visual context, or interpreting temporal references like "this one" by tracking gestures and scene changes. Second, proactive interaction means the model monitors the environment continuously and can speak unprompted when it detects a relevant change, such as spotting an object the user asked to be reminded about. Third, the model manages conversational timing natively, distinguishing background chatter from the primary speaker and deciding when to interject without relying on an external voice activity detector.
ByteDance published end-to-end human evaluation results showing SeedRealtime cuts conversational pacing failures by roughly half compared to cascaded pipelines. The company reports fewer instances of the model being cut off mid-sentence, responding sluggishly after pauses, or being falsely triggered by background noise.
The launch follows OpenAI's GPT-Live announcement in late July and places ByteDance directly in competition with other full-duplex voice efforts from OpenAI, Microsoft, and Google. SeedRealtime's immediate consumer deployment through Douyin gives it a distribution advantage that lab-stage competitors currently lack.
ByteDance also operates Volcano Engine, its cloud AI platform, which suggests an API path for enterprise use, though the company has not yet announced developer access details for SeedRealtime.
The release marks a shift in the voice AI landscape away from turn-based interaction and toward models that watch, listen, and speak simultaneously -- a technical threshold that multiple labs are now crossing within the same quarter.
