Digital assistants have to respond quickly, but correctly—so researchers are studying how real humans navigate that trade-off, to design better machines.

语音助手必须快速但正确地对主人做出回复，因此研究人员正在研究真正的人类如何进行这种权衡，从而设计出更好的机器。

撰文播音：克里斯托弗·因塔利亚塔（Christopher Intagliata）

翻译：陈美娟

校对：张艺箫

If you've ever used a digital assistant like Siri or Alexa, you know the back and forth doesn't quite have the same rhythm as real human-to-human conversation. The pauses are just a little too long. 

如果你使用过像Siri或Alexa这样的语音助手，你就会发现他们对话的前后节奏和真正的人与人之间的对话不太一样。他们停顿的时间长一点。

"We feel this sort of awkward silence building up." Sebastian Loth, a research psychologist at the University of Bielefeld in Germany. "If I ask you something and you just don't respond, it feels like, oh my god this silence is almost killing the room and you can feel it literally building up. So we try to avoid that. And that's the kind of effect that you're seeing with Siri taking a second to respond, you're kind of feeling odd about it." 

德国比勒菲尔德大学的心理学家塞巴斯蒂安·洛斯（Sebastian Loth）说：“我们感到这种尴尬的沉默正在形成。如果我问你一件事而你就是不回答，我会有就要窒息在房间里的感觉，并且会真实地感觉到尴尬的氛围在加深。所以我们尽量避免这种情况。这就是Siri需要一秒钟来回应对话给人的感觉，有点奇怪。”

To study how humans are able to have such fluent, sometimes even overlapping, conversations, Loth and his team set up what's called a "Ghost in the Machine" experiment, in a barroom situation. Real human customers bellied up to a bar, where a robot bartender was waiting. The robot was actually controlled by human operators behind the scenes, who could see and hear the customers through the robot's eyes and ears. 

为了研究人类是怎样进行如此流畅、有时甚至是双方同时进行的对话，洛斯和他的团队在酒吧里做了一个名为“机器中的幽灵”的实验。真人顾客蜂拥而至，酒吧里有一个机器人侍者在那里等着。机器人实际上是由台后的操作人员控制的，操作人员可以通过机器人的眼睛和耳朵看到和听到顾客的举动。

Then Loth and his team observed how the human operators behaved during a couple hundred orders. They found that when the customers started a phrase with "What," the human operators quickly triggered the robot to repeat the offerings of the bar, like "We have coke, orange juice and water," rather than waiting for customers to complete the sentence. But if the customers began a sentence with "I'd like" or "I want," the human operators tended to hesitate, to listen for what came next, rather than acting quickly—and perhaps incorrectly.

洛斯和他的团队观察了人类操作员在几百个命令中的行为。他们发现，当顾客开始说到“什么”时，人类操作员迅速触发机器人语音键钮，让机器人重复酒吧提供的东西，比如“我们有可乐、橙汁和水”，而不是等待顾客说完。但是，如果客户以“我想”或“我想要”开头，那么人工操作员倾向于犹豫，倾听接下来会发生什么，而不是迅速采取行动——不然也许会做错。

"What we found was that they distinguish between the type of request. And more specifically by the error or cost of that. If I misunderstand you and give you the wrong drink that is actually quite an awkward situation. I have to apologize, take the drink away from you, and replace it with the appropriate one. So it's a lot of action and maneuvering involved and it's quite embarrassing if the bartender gets it wrong." The details are in the journal PLOS ONE. [Sebastian Loth et al., Confidence in uncertainty: Error cost and commitment in early speech hypotheses]

“我们发现他们注意区分客户需求的类型。特别是那些和错误或成本有关的需求。如果我误解了你的意思，给了你错的饮料，那其实是很尴尬的情况。我得道歉，把饮料拿走，然后换成正确的饮料。所以这涉及到很多行为和策略，如果调酒师弄错了，那会很尴尬。”该研究的更多细节发表在“PLOS ONE”杂志上。

Loth says cataloguing interactions like these might help future robots better navigate the continuum between certainty and time: to be able to act quickly on limited evidence, to provide speedy service, but without getting things wrong so often that it annoys the user. 

洛斯说，记录这些交互能帮助未来的机器人更好地应对准确性和及时性之间的连贯性：使它们能够根据有限的信息迅速采取行动，快速提供服务，但不会因为频繁出错惹恼用户。

And no, he doesn't really envision replacing human bartenders with robots. "What you'll probably see is that stuff like that will be incorporated in ticket machines, or in booking systems, or in request systems like Siri where you're asking for a route to be displayed on the phone."

当然，他不是真的想用机器人代替人类调酒师。“您更可能会看到的是，这样的机器会被并入售票机、预订系统或像Siri这样你能要求在电话上显示路线的需求系统。”

After all, he says, a real human bartender does a lot more than serve drinks. 

毕竟，洛斯说，真人酒吧招待为顾客提供的服务不只是倒酒。