Google Duplex 用 AI 打电话订餐

2018 年 5 月，Google I/O 大会的一个演示片段让全场安静了下来。台上播放了一段通话录音：AI 打电话给理发店，用自然的语气询问预约时间，中间还带着「嗯」「让我看看」这样的口语停顿，店员全程没有察觉对面不是真人。这个 AI 叫 Duplex，那一刻，「AI 替人打电话」第一次不再是科幻片。

要理解这次演示的分量，得先看 2018 年的语音助手处境。彼时助手们擅长「听懂指令、给出答案」，但一旦进入真正的对话——对方会反问、会改期、会说「我们周四下午不行」——机器就露馅了。Duplex 解决的就是这个最难的环节：在开放、口语化、充满意外的真实通话里维持流畅对话。

演示里最让人印象深刻的其实是那些「人味」细节。语气词、停顿、犹豫，这些通常被视为语音合成的缺陷，Duplex 反而把它们用成了让对话更可信的工具。技术与设计的这种结合，正是它显得逼真的原因。这种拟真也立刻引发了讨论：当机器听起来和人一模一样，它该不该先表明身份？

Google 的回应很实际：会在通话开始时说明这是 AI 代约，并先把 Duplex 限定在预订类场景。这个折中方案本身也成为行业惯例的雏形——后来的电话机器人、客服 AI 大多沿用了「先亮明 AI 身份」的做法。从技术到伦理，Duplex 都成了探路者。

回看 Duplex，它证明了一件事：语音 AI 的价值不只是「听懂」，更是「办成事」。从 Siri 的应答到 Duplex 的办事，语音交互完成了一次能力跃迁。今天全双工语音模型能更自然地和用户聊天，追根溯源，2018 年那段订餐录音是绕不开的起点。

In May 2018 a clip at Google I/O silenced the room. It played a phone call: an AI called a hair salon, asked about appointments in a natural voice, dropping in "um" and "let me see" like a real person. The receptionist never suspected the other end wasn't human. The AI was called Duplex, and in that moment "AI making phone calls for you" stopped being science fiction.

To understand the weight of the demo, look at where voice assistants stood in 2018. They were good at "understand a command, give an answer". But in true dialogue—when the other person pushes back, reschedules, says "Thursday afternoon doesn't work"—machines fell apart. Duplex attacked the hardest part: holding fluent conversation through real, colloquial, unpredictable phone calls.

The most striking detail was the human texture. Filler words, pauses, hesitation—usually treated as flaws of speech synthesis, Duplex turned them into tools for sounding believable. That fusion of technology and design is what made it convincing. And that very believability raised the question: when a machine sounds exactly like a person, should it say so?

Google's answer was practical: the AI would identify itself at the start of calls, and Duplex would be limited to reservation scenarios. That compromise became a template for the industry—phone bots and customer-service AI largely adopted "state you're an AI first". Technically and ethically, Duplex was the pathfinder.

Looking back, Duplex proved that the value of speech AI isn't just understanding—it's getting things done. From Siri's answering to Duplex's doing, voice interaction crossed a threshold. Today's full-duplex voice models hold even more natural conversations, and that 2018 booking call is a start line they all trace back to.