OpenAI 发布 GPT-Live 全双工语音模型

2026 年 7 月 8 日，OpenAI 发布 GPT-Live-1 与 GPT-Live-1 mini。两款都是全双工语音模型：可以一边听一边说。从 2024 年 7 月的 Advanced Voice Mode 算起，ChatGPT 的语音助手第一次不再按"你一句、我一句"的回合协议工作。付费用户的语音默认由 GPT-Live-1 驱动，免费用户使用 mini 版本，iOS、Android 与网页端逐步推送。

"边听边说"听起来像体验细节，实则是交互协议的变化。回合制语音里，打断是一种需要专门处理的异常：模型正在输出，用户开口，系统要么忽略要么硬切。全双工下，打断变成正常操作——模型可以在被插话时停止、等待、再继续，也可以在用户停顿的间隙主动垫一句话而不显得机械。对客服、翻译、驾驶这类实时场景，这种能力比更长的上下文窗口更早改变体验边界。

工程侧的数字说明这条路径并不轻松。官方博客披露，为做到亚秒级响应，团队重新设计了状态化推理、动态上下文管理与异步委派；媒体前端用 Go 重写；WARP 协议把启动阶段的网络往返从 6 次减到 1 次。复杂任务可以后台委派给 GPT-5.5 等模型，回答可以携带可视化卡片。换句话说，这次发布背后是整条语音链路的重做，而不只是模型参数的变化。

官方公布了盲测偏好率高于旧版语音模式的结果——这是发布方数据，值得存疑地读。真正可核验的部分是产品行为：7 月 31 日起，生成的音频会加入 SynthID 水印，把"这段声音是 AI 合成的"变成可机器识别的元数据。在各国合成媒体标识规则陆续落地的当口，这一条让语音产品提前进入合规队列。

全双工语音对行业的意义，类似当年多模态对聊天框的意义：交互的自然度不再是锦上添花，而是产品能否进入真实场景的前提。对开发者来说，构建语音 Agent 的协议栈从"回合"转向"流式"，延迟与打断处理成为新的技术考题。对普通用户来说，变化更直接——和 AI 说话，终于可以像和人说话一样，随时插嘴。

On July 8, 2026, OpenAI released GPT-Live-1 and GPT-Live-1 mini. Both are full-duplex voice models: they can listen and speak at the same time. Since Advanced Voice Mode arrived in July 2024, ChatGPT's voice assistant has worked under a turn-based protocol—you speak, then I speak. These models end that. GPT-Live-1 drives voice by default for paid users, the mini variant for free users, with a gradual rollout across iOS, Android, and the web.

"Listening while speaking" sounds like a product detail; it is a change in interaction protocol. Under turn-based voice, interruption is an exception to handle: the model is talking, the user speaks, and the system either cuts off or fights it. Full-duplex makes interruption ordinary—the model can stop when spoken over, wait, continue, and can offer a natural backchannel during a pause without feeling mechanical. For real-time settings like customer service, interpretation, and driving, this changes the experience boundary earlier than a longer context window would.

The engineering numbers show the cost of this path. For sub-second response, the team redesigned stateful reasoning, dynamic context management, and asynchronous delegation; the media frontend was rewritten in Go; and the WARP protocol cut startup network round trips from six to one. Complex tasks can be delegated in the background to models such as GPT-5.5, and answers can include visual cards. This release is a rebuild of the voice pipeline, not a parameter bump.

OpenAI reports that blind-test preference exceeded the previous voice mode—vendor data, worth reading skeptically. The verifiable part is product behavior: starting July 31, generated audio carries a SynthID watermark, making "this voice is synthetic" machine-detectable metadata. With synthetic-media labeling rules arriving in multiple jurisdictions, this line moves voice products into the compliance queue ahead of schedule.

Full-duplex voice matters to the industry the way multimodality mattered to chat: interaction naturalness stops being decoration and becomes a precondition for entering real scenarios. For developers, the voice-agent stack shifts from turns to streams, and latency and interruption handling become the new engineering test. For ordinary users, the change is simpler—talking to an AI now feels like talking to a person: you can interrupt whenever you like.