亚马逊、苹果、谷歌等公司正在争先后恐地投入数十亿美元,以使语音识别技术成为人类与互联网交互的主要途径。这也将是史蒂夫·乔布斯发布iPhone以来最重要的技术革命。
短短四年前,亚马逊还只是一家很成功的在线零售商,也是美国商用在线主机服务的主要供应商。此外它有也自己的消费电子产品,即人们熟知的Kindle电子书。Kindle虽然是一款大胆的作品,但考虑到亚马逊本身就是卖书起家,这一尝试自然是可以理解的。现在,亚马逊的Echo智能音箱和它的Alexa语音识别引擎又走进了很多家庭,可以说亚马逊在个人计算与通讯领域,已经掀起了自史蒂夫·乔布斯发布iPhone以来的最重要的技术革命。
一开始,它只不过是个看似新奇的小玩意儿。2014年11月,亚马逊发布了Echo智能音箱,它使用了人工智能技术来倾听人类的提问。Echo会扫描联网数据库中的数百万个单词,不论你提出的问题是深邃还是浅显,它都能给出答案。目前,Echo智能音箱的销量已达到4700多万台,其用户来自从阿尔巴尼亚到赞比亚的80多个国家,其服务器每天要回答用户的1.3亿多个问题。亚马逊的语音识别引擎Alexa得名于亚历山大港的古埃及图书馆,它可以按照用户的要求播放音乐,提供天气预报信息或体育比赛的得分,甚至可以远程调节用户家里的室温。它还会讲笑话,回答一些琐碎的问题,抖个机灵,或者开些无伤大雅的玩笑。(比如你可以让它放个屁来听听)。
亚马逊并没有“发明”语音识别技术,实际上语音识别技术已经发明出来几十年了。亚马逊甚至并不是第一家提供主流语音识别应用的科技巨头。苹果的Siri和谷歌语音助手的上市时间要比它早得多。微软Cortana的发布基本上与Alexa在同一时期。但是随着Echo的广泛成功,语音识别领域的竞争骤然激烈了起来,各大科技厂商纷纷投下重注,试图将这些“智能”家居设备变得跟PC甚至和智能手机一样重要。正如谷歌的搜索引擎算法彻底改变了人们的信息消费模式,进而颠覆了整个广告行业一样,由人工智能技术驱动的语音识别技术也会推动类似的革命。亚马逊Alexa部门的首席科学家罗希特·普拉萨德表示:“我们想抹平用户使用互联网时的不顺畅,而最自然的方法就是声音。Alexa不是那种一下子给你展示很多搜索结果,然后说‘选一个吧’的那种搜索引擎,而是会直接告诉你答案。”
各大科技厂商纷纷将人工智能与语音识别技术相结合,其目的远远不只是为了推出一款圣诞购物季最热卖的小家电这么简单。目前,谷歌、苹果、Facebook和微软等公司纷纷砸下重金研发竞品。据投资公司Loup Ventures的分析师吉恩·蒙斯特估算,上述几家科技巨头每年在语音识别技术上的研发支出合计超过了50亿美元,约占年度研发预算总额的10%。他认为,语音识别技术的出现是计算领域的一个“具有重大意义的变化”。他认为,语音指令很快将取代键盘和触屏,成为“我们与互联网交互的最常见的方式”。
随着各大厂商纷纷投入重注,语音识别助手领域的竞争也变得愈发激烈。从研究公司Canalys提供的数据看,目前亚马逊在这一领域暂时领先,它在全球联网音箱市场上的份额达到了42%。谷歌的Home智能家居设备以34%的份额暂居亚军,它搭载了谷歌自研的谷歌助手,据说近期的销量已经反超了亚马逊。苹果的HomePod价格最贵,加入战局也是最晚,虽然市场占有率排名第三,但份额仍远远不如前面两家。去年10月,Facebook也推出了自己的Portal系列影音设备,它们也具备部分语音识别功能。尤其值得注意的是,它搭载的也是亚马逊的Alexa语音识别引擎。
当前,联网智能音箱以及类似设备的市场规模已然不小,而且还在继续增长。不过对于这些科技巨头来说,语音识别技术的价值远远超过这些设备本身。据市场研究机构全球市场观察公司(Global Market Insights)的研究,2017年,全球智能音箱市场的销售额是45亿元,预计到2024年将增长至300亿美元。不过这几家科技巨头显然并不在乎卖硬件的这点小钱,比如亚马逊基本是在将Echo保本甚至亏本销售。在去年欧美地区的假日购物季期间,亚马逊推出了迷你版的Echo Dot音箱,售价只有29美元,ABI研究公司认为这个价格甚至还要低于它的零部件成本。各大厂商之所以肯做赔本生意,就是为了把用户锁定在他们的其它产品和服务上。比如亚马逊就是要通过Echo产品提高亚马逊Prime订阅服务的价值。谷歌则寄希望于语音搜索功能能够引来更多的广告收入。苹果则希望以语音识别技术为工具,将手机、电脑、电视遥控器甚至是车载软件整合在一块,打造一体化的体验。
由于语音识别领域已经吸引了这么多的投资,而且还在快速创新,因此现在预测谁是赢家还为时过早。但有一点大家已经形成了共识,那就是有了人工智能加成的语音识别技术,必然将向今天的智能手机一样,成为我们访问互联网的新用户界面。另外,语音识别技术也将降低人们使用科技的门槛,促进科技的普及。谷歌公司负责谷歌助手与搜索业务的产品与设计的副总裁尼克·福克斯表示:“它让那些不太识字的人也能使用这个系统。另外,人们在开车的时候也可以使用它,做饭的时候也可以用它来听菜谱。每过一段时间,科技就会发生一次结构性的转变。我们认为,语音识别就是这样一种转变。”
虽然如此,但今天的语音识别技术仍然处于比较早期的阶段。它的应用还比较初级,而且它也有一些比较大的风险因素。比如科技公司会不会利用它对用户进行窃听,以及科技公司通过收集公民的语音数据又攫取了多少权力,人们对这些问题都存在着合理的担忧。华盛顿大学电气工程学教授、世界顶级的语音和语言技术科学家玛丽·奥斯坦多夫表示:“有了人工智能语音识别技术,我们就好比从螺旋桨飞机进入了喷气式飞机时代。”她指出,现在的语音识别技术已经能够很好地回答那些直截了当的问题,但在真实语境的对话中,表现得仍然令人失望。“在能识别多少个单词、听懂多少个指令上,人工智能语音识别技术表现得非常出色。但我们毕竟还没进入火箭时代。”
几十年来,科技行业一直坚信,语音识别技术必将成为下一个“杀手级应用”。早在上世纪50年代,贝尔实验室就开发了一个名为奥黛丽(Audrey)的系统,它可以识别从1到9的语音数字。20世纪90年代时已经有了一款名叫Dragon NaturallySpeaking的PC软件,它可以实现简单的语音识别功能,而不需要说话者每说完一个单词就尴尬地停顿一会儿。但直到苹果公司2010年在iPhone上发布了Siri语音助手,消费者才意识到一个拥有强大计算能力的语音识别引擎能做哪些事。大约就在同一时间段,亚马逊这样一家充满了《星际迷航》式幻想的公司(它的老板杰夫·贝佐斯也是一个正牌《星际》迷)开始畅想,能不能将企业号星际飞船上的那种会说话的电脑变成现实。亚马逊公司的普拉萨德曾发表过上百篇关于语音识别人工智能及相关话题的科学文章,他表示:“在我们的畅想中,未来你可以通过语音与任何服务交互。”而Alexa就是为此而生的。它是一台多才多艺的设备,可以让消费者更容易地与亚马逊进行交互。
随着语音识别技术的进步——也就是计算速度越来越快,价格越来越便宜,越来越普及,因此日益主流化——亚马逊、谷歌、苹果等科技厂商也得以更容易地建立一个无缝的网络,利用语音识别技术,将智能家居设备与他们旗下的其他系统连接起来。比如苹果CarPlay的用户下班路上可以告诉Siri,别忘了在苹果电视上下载最新一集的《权力的游戏》,然后让HomePod等我一回家就开始播放。两年前,谷歌也发布了基于语音识别技术的智能家居产品Home,它将谷歌的音乐服务(YouTube)和最新款的Pixel系列手机和平板产品结合在了一起。换言之,每个科技巨头都将语音识别技术当作了连接其多个数码产品的纽带。
上述几个科技巨头个个都有超强的盈利能力,因此他们都有充足的资金来搞研究和营销,最终拿出的产品也各不相同。苹果和谷歌都有自己的移动操作系统,也就是说,iPhone和所有的安卓手机在出厂时就已预装了Siri或谷歌助手。相比之下,亚马逊就得说服用户将Alexa应用下载到他们的iPhone或安卓手机上了。前华尔街分析师蒙斯特认为:“要打开Alexa语音识别应用,就要比Siri和谷歌助手多花一步,这对亚马逊是一个明显的劣势。” 而相比之下,Siri和谷歌助手只需用户喊一声它们的名字就能激活。
不过,iOS和Android是面向所有第三方开发者的,而Alexa应用同时兼容这两个平台,也就是说,两个平台上的开发者都可以写Alexa的程序。亚马逊CEO杰夫·贝佐斯今年早些时候曾在一次财报发布会上称:“有来自150多个国家的数万名开发者”都在构建Alexa的应用程序,并将它们集成到非亚马逊的设备里。而合作伙伴也是各大语音识别应用竞争的一个竞争战场。现在,Sonos公司的“电声棒”、Jabra公司的耳机,以及宝马、福特、丰田等公司的汽车都已用上了Alexa。谷歌的语音识别程序则被集成到了索尼、铂傲的音响、August公司的智能门锁和飞利浦的LED照明系统上。苹果的HomPod则与First Alert公司的安全防卫系统和霍尼韦尔公司的智能恒温器进行了合作。谷歌副总裁尼克斯表示:“这些合作的好处是将语音识别功能整合到了整个智能家居生态系统,我不用打开手机也能使用应用程序了。我只要说一声:‘让我看看谁在门口’,门前的监控视频就会自动显示出来。总之,它通过统一实现了简化。”
人工智能一直是反乌托邦文化里的常客,特别是在《终结者》和《黑客帝国》系列电影里,智能机器人甚至造了人类的反,将人类逼到了“亡球灭种”的边缘。不过庆幸的是,现在的我们离被机器人奴役还有很远。不过人工智能技术的进步,以及廉价计算设备的普及,已经让很多具有科幻感的构思成为了现实。早期的语音识别程序虽然也不错,但也没有超过编写它们的程序员的最高水平。但现在这些应用却变得越来越好了,这是因为它们通过互联网与数据中心连接,而且科技公司花了好几年时间,用大量数据对这些算法进行“训练”,使其学会了识别不同的语言模式。现在,这些人工智能语音识别应用不仅能识别单词、方言和俗语,甚至还能根据上下文分析语义(比如通过分析呼叫中心的客服代表与客户的电话录音,或者分析用户与数字助手的互动)。
The current market for connected speakers and similar gadgets is big and growing—but not necessarily the most dramatic voice-related opportunity for the tech titans. Global Market Insights, a research firm, pegs global 2017 smart-speaker sales at $4.5 billion, a number it projects will grow to $30 billion by 2024. The hardware revenues, however, are largely beside the point. Amazon, for example, has sold the Echo at breakeven or less. Last holiday season it offered the bare-bones Echo Dot for $29, which ABI Research reckons is less than the cost of the device’s parts. Instead, each major player has a strategy that in some way feeds its larger goal of locking in customers to its other goods and services. Amazon, for one, uses the Echo line to increase the value of its Amazon Prime subscription service. Google hopes voice searches will eventually boost the already massive trove of data that feeds its advertising franchise. With Siri, Apple sees a way to tie together its phones, computers, TV controllers, and even the software that automakers are tying into their onboard systems.
It’s too soon to predict a winner, what with all the investment and fast-moving innovations. But it’s safe to say the industry has coalesced around the notion that voice technology, enhanced by recent advancements in artificial intelligence, is the user interface of tomorrow. And it promises to have a democratizing impact on an industry that has separated novices from experts. “Voice enables all kinds of things,” says Nick Fox, a Google vice president who oversees product and design for the Google Assistant and Search. “It enables people who are less literate to use the system. It enables people who are driving. It enables people while cooking to hear a recipe. Every once in a while there is a tectonic shift in technology, and we think voice is one of those.”
For all that, voice recognition remains in its infancy. Its applications are rudimentary compared with where researchers expect them to go, and there’s a significant ick factor associated with voice. Legitimate concerns linger as to how much the tech companies are eavesdropping on their customers—and how much power they are accumulating in the form of data derived from the spoken information they are collecting. “With A.I. voice recognition, we’ve gone from the age of the biplane to the age of the jet plane,” says Mari Ostendorf, a professor of electrical engineering at the University of Washington and one of the world’s top scientists on speech and language technology. She notes that computers have gotten good at answering straightforward questions but still are relatively hopeless when it comes to actual dialogue. “It’s truly impressive what Big Tech has done in terms of how many words voice A.I. can now recognize and the number of commands it can understand. But we’re not in the rocket era yet.”
VOICE RECOGNITION HAS BEEN the next killer app for decades. In the 1950s, Bell Labs created a system called Audrey that could recognize the spoken digits one through nine. In the 1990s, PC users installed Dragon NaturallySpeaking, a program that could process simple speech without the speaker having to pause awkwardly after each word. But it wasn’t until Apple unleashed Siri on the iPhone in 2010 that consumers got a sense of what a voice-recognition engine tied to massive computing power could accomplish. Around the same time, Amazon, a company full of Star Trek aficionados—and led by a true Trekkie in CEO Jeff Bezos—began dreaming about replicating the talking computer aboard the Starship Enterprise. “We imagined a future where you could interact with any service through voice,” says Amazon’s Prasad, who has published more than 100 scientific articles on conversational A.I. and other topics. The result was Alexa, a multifaceted device designed to let consumers communicate more easily with Amazon.
As voice recognition improves—which it does as computing power gets faster, cheaper, more ubiquitous, and thus more mainstream—Amazon, Google, Apple, and others can more easily build a seamless network where voice links their smart home devices with other systems. It’s possible for Apple CarPlay users, for example, to tell Siri on the drive home to slot the latest episode of Game of Thrones as “up next” on their Apple TV and to command their HomePod to play it once they’ve arrived. Two years ago, Google released its voice-enabled Home that ties together its music offerings, YouTube, and its latest Pixel phones and tablets. Each tech giant, in other words, sees voice as a tether to the myriad digital products it is creating.
The combatants, each wildly profitable and therefore able to fund ample research and marketing efforts, bring different assets to the table. Apple and Google, for example, own the two dominant mobile operating systems, iOS and Android, respectively. That means Siri and Google Assistant come preinstalled on nearly all new phones. Amazon, in contrast, needs to get consumers to install and then open the Alexa app on their iPhones or Android devices. “The extra step to open the Alexa voice app puts Amazon at a distinct disadvantage,” says Loup’s Munster, formerly a Wall Street analyst of computer companies. By contrast, all that’s required to activate Siri and the Google Assistant is to say their names.
That said, iOS and Android are open to third-party developers of all stripes, and Amazon is one of them¬—meaning that nothing is stopping developers on both platforms from writing Alexa programs. Bezos bragged in an earnings release earlier this year that “tens of thousands of developers across more than 150 countries” are building Alexa apps and incorporating them into non-Amazon devices. Indeed, partnerships are a key battleground for voice applications. Alexa is built into “soundbars” from Sonos, headphones from Jabra, and cars from BMW, Ford, and Toyota. Google boasts integrations with audio equipment makers Sony and Bang & Olufsen, August smart locks, and Philips LED lighting systems, and Apple has partnerships that allow its HomePod to work with First Alert Security systems and Honeywell smart thermostats. “The beauty of these partnerships,” says Google’s Fox, “is that they allow us to link voice into the whole smart-appliance ecosystem. I don’t have to open my phone and go to an app. I can just say to the device, ‘Show me who’s at my front door,’ and it will pop right up. It’s simplifying by unifying.”
Artificial intelligence has long been a staple of dystopian popular culture, notably from films such as The Terminator and The Matrix, where wickedly clever machines rise up and pose a threat to humankind. Thankfully, we’re not there yet, but advances in A.I. and the availability of cheap computing have made impressively futuristic applications a reality. Early voice-recognition programs were only as good as the programmers who wrote them. Now these apps keep getting better because they are connected through the Internet to data centers. These complex mathematical models sift through huge amounts of data that companies have spent years compiling and learn to recognize different speech patterns. They can recognize vocabulary, regional accents, colloquialisms, and the context of conversations by analyzing, for example, recordings of call-center agents talking with customers or interactions with a digital assistant.
图片来源:Pope: Heinz-Dieter Falkenstein—Getty images; Edison: Bettmann/Getty Images; Audrey: Courtesy of Nokia Bell Labs: Telephone: Sheila Terry—Science Source; Shoebox: Courtesy of IBM Corporate Archives, © 1961 Corporation; HAL: Kevin Bray—MGM/Photofest; Harpy: Raj Reddy—Youtube; Devices: Courtesy of Amazon, Apple, and Google