There have been a number of incidents in which Alexa's digital assistant, like the Amazon, has misinterpreted something that was overheard and it began to randomly record private individuals' records To send conversations or audio files that were kept by the company in the wrong hands. But a real stranger could potentially hear everything you tell Alexa from the start with a report in Bloomberg on Wednesday.
According to Bloomberg's report, Amazon employs "thousands" of people around the world whose job is to improve Alexa's voice. Recognition functions. This team has access to voice recordings of real-world customers using Alexa-powered devices both at home and at work (the report only mentions echo speakers directly, although Alexa also runs on cell phones and many third-party devices). These shots are "transcribed, annotated, and then reintegrated into the software," Bloomberg wrote as part of an effort to further enhance Alex's ability to recognize speech without human intervention.
The process is necessary because Alexa sets limits on his ability to train himself, especially when it comes to garbled phrases, accents, slang, regional words or other languages. Last year, Wired reported that "active learning techniques" in which the system identifies areas where it could be improved by human support, "have helped to significantly reduce Alexa's error rates." trying to apply the previously learned skills to new ones has helped developers "reduce the work they would otherwise face".
New is "Self-Learning," in which Alexa tries to understand contextual hints to understand commands that are not hyperspecifically output (eg, "Alexa, play 102.5 FM The Bone" or "Alexa, play the bone "). According to Wired, Amazon plans to persuade Alexa to recognize the emotions of users, which according to critics could lead to manipulative marketing strategies. In an article in Scientific American last month, Amazon's director of applied science, Ruhi Sarikaya, argued that such huge amounts of data will soon require an analysis that requires speech recognition systems to move from a "supervised" learning model to a semi-supervised, low-level supervised and unattended learning. Our systems need to learn how to improve.
Bloomberg interviewed seven different sources about the program, some of which said Amazon workers analyze about 1,000 audio clips per nine-hour shift. Most of the time was "banal," wrote Bloomberg:
A Boston laborer said he had accumulated accumulated voice data for certain utterances such as "Taylor Swift" and commented on it to suggest that the seeker meant the musician. Occasionally the listeners pick up things that echo owners would rather keep private: a woman who sings badly in the shower, say, or a kid crying out for help. The teams use internal chat rooms to exchange files when they need help parsing a confused word – or encounter an amusing recording.
On other occasions, however, workers have heard what they took to be crimes, including those who considered them sexually assaulted. Amazon told the workers in Romania that it was not the company's job to intervene, Bloomberg wrote. Others told the news agency that every auditor sees up to 100 shots a day, when Alexa seems unaware of being activated by a user with a wake-up word or command (for example, pressing a key).
Amazon characterizes the number of records that are actually analyzed by humans as an "extremely small sample" in a statement to Bloomberg, adding that it served exclusively for the purpose of "[improving] the customer experience". In addition, the process was described as low risk:
We have strict technical and operational guarantees and have a zero tolerance policy for the abuse of our system. Employees do not have direct access to information that can identify the person or account as part of this workflow. All information is treated with high confidentiality and we use multi-factor authentication to restrict access, service encryption and monitoring of our control environment for protection.
Bloomberg noted, however, that a screenshot provided by an examiner "shows that the recordings were sent. Alexa reviewers do not provide a user's full name and full address, but are provided with an account number and the user's first name and the serial number of the device.
According to Bloomberg, an Apple whitepaper, the Siri Voice Assistant only writes people to analyze recordings that contain "no personally identifiable information and are stored for six months at random. "The records can be removed later for long-term storage of random IDs. Google Auditors can only access distorted audio.