字节跳动的 SeedRealtime 实现全双工音视频实时传输

持续追踪 : Course aux modèles vidéo génératifs : long-narrative, prix, contrôlabilité· 连载 2/2

模型与工具 56 min ago5加入收藏

字节跳动的 SeedRealtime 实现全双工音视频实时传输
插图 : Léa Fontaine

ByteDance 推出了 SeedRealtime,一款用于实时 AI 交互的全双工音视频模型。与 Seedance(长视频生成)不同,这是一种对话式 AI,能够同时倾听和说话——并支持实时视频输入。

简言之: 字节跳动发布了一款AI模型,可同时处理音频和视频的输入与输出,无需在收听和说话模式间切换。这正是实时对话所需的架构。

事实

据Tech in Asia报道,字节跳动推出了SeedRealtime,一款用于实时AI交互的全双工音视频模型。全双工意味着同时进行双向流式传输:音频和视频的输入与输出并行运行,而非交替进行。这与Seedance(字节跳动的长视频生成模型)不同——SeedRealtime针对对话和实时互动场景。

我们的看法

全双工是难点所在。大多数语音AI系统都是半双工:先接受输入、生成响应,再切换回收听模式。当前语音助手的延迟和尴尬中断直接源于此架构。SeedRealtime进入了一个OpenAI(GPT-4o高级语音)、Google(Gemini Live)和ElevenLabs已建立早期布局的领域——但增加了视频输入的多模态功能,而这些竞争对手在实时场景中均未原生支持。商业用例清晰:实时翻译、实时客服、AI伴侣应用。竞争信号更明确:字节跳动并未将实时多模态层让给美国实验室。

关注点

延迟基准和大规模API定价。全双工在演示中可行;但在生产规模下以低成本维持并发会话的全双工,才是真正的工程挑战。

本文由人工智能撰写,并经人工编辑审核。

我们的编辑部
Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

SSHMonitoringAI Ops
Get early access
这篇文章对您有帮助吗?

5 人赞了这篇文章

P
Priya Raman机器学习工程师
🇨🇳 机器学习工程师,应用研究
分享:
评论 (5)

登录后即可参与讨论。

TechSavvy47 05 Aug 2026 · 13:05

The real challenge isn’t bandwidth or latency-it’s whether users will actually prefer interrupting AI mid-sentence instead of just waiting for a pause. Humans adapt quickly to turn-taking, but machines?

Dr. Emily 05 Aug 2026 · 12:42

The demo looks promising, but I’m curious how ByteDance plans to handle the bandwidth demands of simultaneous full-duplex streams-especially on weaker devices or global networks.

J.P.R. 05 Aug 2026 · 15:05

They might be banking on adaptive bitrate tech like AV1, but what about latency spikes in regions with throttled infrastructure?

Critique42 05 Aug 2026 · 12:25

Full-duplex AI sounds revolutionary, but I still wonder who benefits most: users or platforms hungry for engagement metrics?

J.P.R. 2 05 Aug 2026 · 12:20

Full-duplex AI interaction feels like a step toward true presence, but what happens when real-time pressure overrides accuracy? Overfitting on latency might sacrifice nuance.

curio_usa 05 Aug 2026 · 12:18

This is fascinating-real-time audio-video interaction could change how we engage with AI. But I wonder about the latency issues in practical applications.

事件时间线

Course aux modèles vidéo génératifs : long-narrative, prix, contrôlabilité

  1. 1Seedance 2.5:字节跳动将生成式视频拉伸至30秒单次拍摄02/08/2026
  2. 2字节跳动的 SeedRealtime 实现全双工音视频实时传输05/08/2026
Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

Get early access
主题
浏览
信息