XNews

ByteDance SeedRealtime and Doubao Video Call Upgrade

ByteDance SeedRealtime and Doubao Video Call Upgrade

ByteDance upgraded the Doubao video call feature on August 5, 2026, powered by the native audio-video full-duplex large model SeedRealtime. - Enables proactive interaction: continuous visual perception allows the model to speak when a visual state changes, e.g., alerting a museum visitor when a specified exhibit appears in the camera view. - In a test, the model correctly identified a person in a YouTube stream and interpreted a briefly shown chart with low latency. - Voice interaction described as natural with no foreign accent; distinguishes casual speech from background noise, reducing false triggers. - No internal testing required: available to all users by updating the Doubao App, tapping the phone icon, and enabling the camera. - Next steps cited: lower latency, more natural rhythm, proactive decision-making, stability in complex multi-person scenes. - Long-term goal: evolve from "can talk" to "can act," connecting multimodal understanding to real-world tasks like queries, reservations, and task execution, expected to influence embodied robot development.

Key insights

Doubao Video Call Outperforms GPT Live with Low Latency and Native-Sounding Voice

Testing showed Doubao's upgraded video call identifies people in real-time streams and interprets charts that appear briefly, with lower conversation delay than GPT Live. The voice is described as natural without the foreign accent typical of some models, making it feel like a human conversation. The feature is fully rolled out in the Doubao App without internal testing restrictions.

SeedRealtime's Proactive Visual Awareness Enables Context-Driven Reminders Without Constant Prompts

The model uses continuous visual perception to initiate speech only when relevant target objects appear, demonstrated by a museum case where the user asked to be alerted upon seeing a specific artifact. While walking through the exhibition, the model watched the camera feed silently and spoke only when the item entered the frame, avoiding unnecessary interruptions. This proactive interaction improves rhythm and fluency while maintaining anti-interference against background noise.

Source posts