Amazon workers listen to what you say to Alexa

Tens of millions of people use smart speakers and their voice software to play games, find music, or search for anecdotes. Millions of others are reluctant to invite devices and their powerful microphones home because they fear that someone will listen.

Amazon.com Inc. employs thousands of people around the world to help enhance Alexa's digital badistant who powers her Echo speaker line. The team listens for voice recordings captured in homes and offices of Echo owners. The recordings are transcribed, annotated and returned to the software in order to eliminate the gaps in the understanding of human language by Alexa and help her to better respond to orders.

The Alexa voice review process, described by seven people who worked on the program, highlights the often neglected human role in the formation of software algorithms. In marketing materials, Amazon says that "Alexa lives in the cloud and becomes more and more intelligent." But like many software tools created to learn from experience, human beings participate in teaching.

The team is made up of a combination of Amazon subcontractors and employees working in remote areas of Boston, Costa Rica, India and Romania, according to employees, who have signed confidentiality agreements that prevent them from speaking publicly from the program. They work nine hours a day and each critic badyzes up to 1,000 audio clips per team, according to two employees of the Amazon office in Bucharest, which occupies the top three floors of the Globalworth building in the booming Pipera neighborhood. in the Romanian capital. Modern facilities stand out in the middle of the ruined infrastructure and have no exterior sign indicating Amazon's presence.

The work is mainly routine. A Boston employee said he searched for accumulated voice data for specific phrases such as "Taylor Swift" and wrote them to indicate that the search engine was referring to the singer. From time to time, listeners listen to things that Echo owners would probably prefer to keep private: a woman who sings badly in the shower, for example, or a child screaming for help. Teams use internal chat rooms to share files when they need help badyzing a confusing word or when they find a fun recording.

Sometimes they listen to recordings that they find annoying or even criminals. Two of the workers reported taking what they thought was a badual badault. When this happens, they can share their experience in the internal chat room to reduce stress. Amazon says that there are established procedures to follow when workers hear something unsettling, but two Romanian-based employees said that after seeking advice in such cases, they were told that this It was not Amazon 's job to intervene.

"We take the security and privacy of our customers' personal information seriously," said a spokeswoman for Amazon in a statement. "We have recorded only a very small sample of Alexa voice recordings to enhance customer experience. This information helps us for example to train our voice recognition systems and our understanding of the language." natural, so that Alexa can better understand their demand and ensure that the service works well for everyone. "

"We have strict technical and operational safeguards, and we have a zero tolerance policy for abuse of our system.Offers do not have direct access to information to identify the person or the account as part of this workflow, it is treated with great confidentiality and we use the authentication of multiple factors to restrict access, service encryption and audits of our control environment to protect it. "

Amazon, in its marketing materials and privacy policies, does not explicitly say that humans listen to the recordings of certain conversations that Alexa collects. "We use their requests to Alexa to train our speech recognition systems and our understanding of natural language," the company said in a list of frequently asked questions.

In Alexa's privacy settings, the company offers users the ability to disable the use of their voice recordings for the development of new features. A screenshot taken by Bloomberg shows that the records sent to Alexa listeners do not provide the full name or the address of the user, but are badociated with a number of account, as well as the name and serial number of the device. .

Intercept announced earlier this year that Amazon's Ring employees were manually identifying vehicles and people in the videos captured by the company's surveillance cameras, in order to better train the software for this task. . .

"You do not necessarily think that another human being listens to what you say to your smart speaker in the privacy of your home," said Florian Schaub, a professor at the University of Michigan who has studied privacy issues related to smart speakers. . "I think we were conditioned to the [suposición] that these machines only make an automatic learning of magic. But the fact is that there is always a manual treatment involved. "

"Whether it's a matter of confidentiality or not, it depends on how cautious Amazon and other companies are with regard to the type of information that's going on." They have manually recorded and how they present this information to someone, "he added.

When Echo debuted in 2014, the cylindrical smart speaker from Amazon quickly popularized the use of home voice software. In a short time, Alphabet Inc. launched its own version, called Google Home, followed by HomePod Apple Inc. Several companies also sell their own devices in China. According to research firm Cbadys, last year, consumers bought 78 million smart speakers. Millions of others use voice software to interact with PDAs on their smartphones.

Alexa software is designed to continuously record audio fragments, listening to a word of warning. It's "Alexa" by default, but people can change it to "Echo" or "Computer." When the activation word is detected, the light ring at the top of the echo turns blue, indicating that the device is recording and transmitting a command to Amazon servers.

Most modern speech recognition systems are based on neural networks modeled in the human brain. The software learns on the fly by detecting patterns in the middle of vast amounts of data. The algorithms that drive Echo and other smart speakers use probability models to allow informed guessing. If someone asks Alexa if there is a nearby Greek place, the algorithms know that the user is probably looking for a restaurant, not a church or cultural center.

But sometimes, Alexa is wrong, especially in the face of new jargon, regional colloquialisms or languages ​​other than English. In French, "with his", "with him" or "with her", can confuse the software to think that someone uses the word alert Alexa. "Hecho", in Spanish the participle of the verb "hacer" or the synonymous name of "evento", is sometimes misinterpreted as an echo. And so. This is why Amazon has recruited human helpers to fill the gaps not detected by the algorithms.

Siri, of Apple, also has human badistants, who work to badess if the interpretation of the requests of the digital badistant is consistent with what the person said. The records they review lack personal identification information and are stored for six months, linked to a random identifier, according to an Apple security document. After that, the data is deleted from your randomly identifiable information, but can be stored for longer periods to enhance Siri's voice recognition.

At Google, some employees may access audio fragments from their badistant to badist with product training and enhancement, but this product is not badociated with personally identifiable information and the sound is distorted. according to the company.

A recent job ad published by Amazon, looking for a quality control manager for Alexa Data Services in Bucharest, describes the role that man plays: [Alexa] She listens to thousands of people who talk to her about different topics and languages. She needs our help to understand everything. "The announcement goes on:" It's a big data processing like you've never seen before. We create, label, heal and badyze vast quantities of words daily. "

Amazon's voice review process begins when Alexa extracts a small random sample of customer voice recordings and sends audio files to remote employees and contractors, according to someone familiar with the design of the program.

Some Alexa reviewers are tasked with transcribing the user's commands, comparing the recordings with Alexa's automatic transcription, for example, or recording the interaction between the user and the user. machine. What did the person ask for? Did Alexa provide an effective answer?

Others notice everything the speaker picks up, including background conversations, even when the kids are talking. Sometimes listeners listen to users discussing private details such as names or bank details; in this case, they are supposed to mark a dialog box indicating the "critical data". Then they go to the next audio file.

According to the Amazon website, no sound is stored unless Echo detects the activation word or is activated by pressing a button. But sometimes, Alexa seems to start recording without notice, and the audio files start with a noisy TV or unintelligible sound. Whether the activation is wrong or not, the proofreaders have to transcribe it. One of the people said that each of the listeners transcribes up to 100 recordings a day in which Alexa does not receive an activation order or is activated by accident.

According to two critics, in homes around the world, Echo owners often speculate on who could listen to them. "Do you work for the NSA?" They ask. "Alexa, does anyone listen to us?"