What a smart speaker is
A smart speaker combines three things in one box: a speaker for playing music and speech, several microphones that listen for a spoken command, and a network connection that sends the recorded request to a remote service and returns a reply. The device itself is simple. The intelligence lies mostly in the software that works out what was said and what to do about it.
Which device was the first is disputed, and the answer is entangled with the history of the internet and home audio. Various speakers with internet streaming and voice features were built earlier in prototypes and niche products. The model that is most often described as the first widely sold one is the Amazon Echo, which the company announced in November 2014, first offering it to a limited group of customers before opening sales more broadly in 2015. That is a narrower claim than “the first,” and it is the one that can be defended.
Speech recognition before the speaker
The ability to recognize speech has a much longer history, and it did not begin with the home.
- 1950s. A Bell Labs system known as Audrey, from 1952, is usually described as recognizing spoken digits from a single speaker, with limited accuracy.
- 1960s. IBM demonstrated a machine called Shoebox in the early 1960s, which responded to a small vocabulary of spoken words, mainly digits and a few commands.
- 1970s. Research funded by the U.S. Defense Advanced Research Projects Agency produced systems such as Harpy at Carnegie Mellon, which could recognize on the order of a thousand words, within constrained grammar.
- 1980s. Statistical approaches, in particular hidden Markov models, replaced many hand-written rules. These handled the natural variation in speech better.
- 1990s. Commercial dictation software for personal computers appeared, and users could talk to their computers, though they often needed to train the program to their voices.
Progress was steady but slow, and until the late 2000s, speech recognition was seen as useful only in narrow cases such as dictation and telephone menus.
From dictation to assistants
Several changes made speech useful for everyday tasks. Increases in computing power allowed larger models. Large collections of recorded speech provided training data. And in the 2010s, machine learning methods based on neural networks sharply improved recognition accuracy. These developments sit within the broader field of artificial intelligence.
Voice assistants arrived on phones first. Apple’s Siri began as a separate app in 2010 and was built into the iPhone in 2011. Google offered voice search and later its own assistant, and Microsoft released Cortana in 2014. Part of the recognition ran on the phone, while the heavy processing ran on remote servers, a design that depended on cloud computing and good network connections.
Putting the assistant in a speaker
A phone assistant has a limitation: someone has to pick up the phone. A device that sits in a room and listens all the time solves that, but it creates technical challenges.
- Far-field recognition. The user may be across the room with a television or music playing. Arrays of several microphones and signal processing help the device pick out a voice from noise and its own sound.
- Wake words. To avoid sending everything to servers, the speaker listens locally for a trigger phrase, and only then records and sends a request. How this works in practice is a matter users have debated.
- Power. The device is plugged in, which removes battery limits that phones face.
The Echo combined these in a cylinder and offered an assistant named Alexa. Google followed with Google Home in 2016, and Apple released the HomePod in 2018. Other makers produced speakers that supported one of the assistants. Over the next few years, speakers with screens, small ones for bedrooms, and speakers built into other appliances became available.
Link to the connected home
Smart speakers also became a hub for the connected house. A speaker can pass commands to lights, plugs and thermostats that support the assistant, so that a person can say “turn off the lights.” This drew on the earlier work on home automation and on smart home networking, and it gave a new type of control that does not need a screen. Music, timers, weather, news and questions became the most common uses in practice, with smart home control used by a smaller share of owners.
Concerns
Always-listening microphones raise questions. Companies have said that devices send audio only after hearing a wake word, but there have been accidental activations, and reports that some recordings were reviewed by human staff to improve accuracy. These stories led to changes in privacy controls, such as options to delete recordings. Whether those controls are enough remains a matter on which people disagree.
What it shows
The smart speaker is a good example of a product that is new in form but old in its parts. Speech recognition took more than half a century to mature, and cloud services, cheap microphones and wireless networks had to be in place as well. The first device to put them together in a form people would buy deserves credit, but the speech research behind it came from decades of work at many labs.