ByteDance’s Seed research group released SeedRealtime, a single model that processes live video, audio, and text together and responds without waiting for a clean break in the conversation. The lab calls this full duplex: the system can listen, watch, and reply at the same time, the way a person on a video call does, instead of trading discrete turns with a user.
That distinction matters because most voice assistants running today, including many enterprise call center bots and earlier consumer products, still rely on a cascaded pipeline. A speech-to-text model transcribes the user, a language model drafts a reply, and a separate text-to-speech model reads it aloud. Each handoff adds latency and strips out tone, pacing, and the ability to interject mid-sentence. OpenAI’s GPT-4o Realtime API and Google’s Gemini Live already replaced that chain with native speech-to-speech models, cutting round-trip delay and smoothing interruptions. Neither product, as offered today, keeps a live camera feed inside that same real-time loop.
SeedRealtime’s contribution is folding continuous video into the native audio loop rather than bolting vision on as a separate call. ByteDance says the model tracks a scene as it changes, a new object appearing or a person walking into frame, and decides on its own whether to speak up, stay quiet, or invoke a tool, instead of only responding when addressed directly. In multi-speaker settings, the lab says the system can tell voices apart, follow who is talking to whom, and ignore a ringing phone or side conversation without losing the thread of the primary task.
ByteDance’s own comparison, run against its earlier cascaded systems rather than a named competitor, claims the new model roughly halves conversational pacing problems and trims false triggers, latency, and interruptions. The company has not published the evaluation size, the scoring method, or which baseline it used, so the “roughly half” figure should be read as ByteDance grading its own homework. No third-party benchmark accompanies the release, and the announcement does not name the specific ByteDance product, such as its Doubao assistant, where SeedRealtime actually runs.
The lab says the model has already reached large-scale deployment, which would make it one of the first audio-visual full-duplex systems operating at consumer scale anywhere, ahead of comparable video-input features still in limited preview at rival labs. That claim rests entirely on ByteDance’s own disclosure. The company gives no user count, no market, and no external confirmation of scale.
ByteDance also laid out where it wants to take the model next: shorter delay between hearing something and responding to it, better handling of interruptions and backchannel noise, more reliable speaker tracking in crowded rooms, and eventually letting the model act on what it sees rather than only narrate it. That last step would move SeedRealtime from a voice interface toward an agent that operates a device on the user’s behalf.
For builders evaluating realtime voice stacks, the near-term test is not ByteDance’s self-reported pacing gains but whether SeedRealtime’s API or SDK reaches developers outside China with documented latency numbers. Until it does, OpenAI’s and Google’s audio-only realtime products remain the only options in this category with independently verifiable performance data, and any team weighing a video-aware assistant should track ByteDance’s next disclosure closely before assuming parity has already arrived.
Published by ByteDance Seed on August 6, 2026.
