火山引擎 11 月 19 日正式上线豆包实时音视频通话模型(SeedRealtime 1.0)API 服务,开发者可将实时多模态交互能力接入应用。该模型把声音、画面、时序与表达统一到同一个端到端模型中,在连续的多模态信息流上实时交互。Volcano Engine officially launched the API for its Doubao SeedRealtime 1.0 audio-video model on Nov 19, letting developers integrate real-time multimodal interaction. The model unifies sound, vision, timing and expression in a single end-to-end model operating on continuous information streams.

相比传统级联系统依赖 ASR、VLM、TTS 多模块串联带来的延迟叠加与信息损耗,SeedRealtime 原生支持声音、画面与时序信息的深度融合,既能结合场景消解同音词歧义,也能理解视觉中的时序指代,并能在合适时机主动开口。火山引擎称其音视频对话节奏问题减少了一半,可用于语音助手、智能硬件、陪伴式应用与实时客服等场景。Compared with cascaded systems relying on ASR, VLM and TTS modules, SeedRealtime natively fuses sound, vision and timing — resolving homophone ambiguity from context and understanding visual temporal references, and can speak proactively at the right moment. Volcano Engine says conversational rhythm issues dropped by half, targeting voice assistants, smart hardware, companion apps and real-time customer service.