ByteDance launches SeedRealtime full-duplex AI model
SeedRealtime’s pitch is fewer awkward handoffs in live multimodal conversations.
ByteDance says the model handles audio, video, and text in one architecture, letting it watch, listen, and respond over continuous streams. It decides when to speak by tracking scenes, speakers, pauses, gestures, and background chatter instead of relying on separate voice-activity rules. The company showed demos in noisy public settings and says human evaluators found half as many audio-visual pacing problems versus cascaded models. ByteDance says it is fully rolled out, but the note gives no API, pricing, regions, or access details. TestingCatalog's note
ByteDance says the model handles audio, video, and text in one architecture, letting it watch, listen, and respond over continuous streams. It decides when to speak by tracking scenes, speakers, pauses, gestures, and background chatter instead of relying on separate voice-activity rules. The company showed demos in noisy public settings and says human evaluators found half as many audio-visual pacing problems versus cascaded models. ByteDance says it is fully rolled out, but the note gives no API, pricing, regions, or access details. TestingCatalog's note
score 7