Amazon Echo

A cylinder that listens

Amazon Echo is a speaker that also works as a voice assistant. On the day it launched it was described as a nine-inch-tall cylinder acting as an artificially intelligent assistant and a high-end speaker in one device^[2]. It was always on, and it responded once it was addressed as "Alexa"^[2]. That single design decision — a device that sits in a room, listening for its name — is the core of the product.

Where it came from

Echo was launched on 6 November 2014 by a small multidisciplinary team at Amazon^[1]. The team's stated aim was to use artificial intelligence to make everyday convenience easier^[1]. One acknowledged inspiration was the computer in Star Trek^[1]: a machine you simply talk to, without a keyboard or a screen.

The launch itself was unusual. The device was sold by invitation only, at USD 199 for customers who were not Prime members and USD 99 for Prime members^[2].

Four problems to solve

Amazon Science lists four foundational AI tasks the device had to handle from the start: wake word detection, automatic speech recognition, natural-language understanding and text-to-speech synthesis^[1]. In plain terms, the system has to notice that someone has said its trigger word; turn the sound that follows into text; work out what the person actually wants; and say something back in synthesised speech. Each of these is a separate research problem, and each can fail on its own.

Not all of that ran on the cylinder itself. According to Amazon's launch site, the brain of Echo ran in the cloud on Amazon Web Services, so the system continually learned and adapted to speech patterns, vocabulary and personal preferences^[2]. Keeping the heavy models in a data centre rather than inside a small speaker is what allowed the product to keep changing after it had shipped: the hardware in a room stays the same, while the software behind it does not.

What improved

Amazon Science reports the trajectory over the following five years. Wake-word and speech recognition errors fell fourfold, and natural-language-understanding errors threefold^[1]. Those two numbers cover different halves of the pipeline: recognising the words correctly, and interpreting them correctly. A fourfold and a threefold reduction mean the same device, with the same microphones, became substantially more reliable at both.

The voice going back out improved too. In MUSHRA listener studies — a standard listening-test method in which people rate audio samples — the naturalness gap between Alexa's speech and human speech fell by 80%^[1]. A smaller gap means synthesised speech that is harder to tell apart from a recording of a person.

Why it matters

Echo took a set of research problems that had mostly been demonstrated in labs and phones and put them into a piece of furniture in a home. The recipe it settled on — an always-on microphone array, a cloud backend, a wake word, and a pipeline running speech recognition, language understanding and speech synthesis — became the standard shape of the consumer voice assistant. Later work on the assistant was reported by Amazon under the name Alexa, the same word users say to wake the device^[1].

The limitations are worth stating plainly. A device that is always on and sends audio to a cloud service raises questions about privacy and about what happens when the network is unavailable. And the four tasks it depends on are statistical: they reduce errors, they do not eliminate them, which is why the error-rate figures above are framed as reductions rather than as solved problems.