The company assured that human intervention is needed only in rare cases and to improve speech recognition.
Amazon’s Voice Assistant from Amazon answers questions and performs user tasks, recognizing their speech. However, the company has a team of people who listen to what the owners say and helps the device to understand their requests correctly. About this Bloomberg told a few anonymous employees.
- According to the publication, several thousand Amazon employees around the world listen to the voice assistant recordings. Audio is transcribed, divided into semantic parts, annotated and returned to the device. – it is necessary to better distinguish speech commands;
- The teams are located in Boston, Costa Rica and Romania: they work nine hours a day and during this time they review up to a thousand audio recordings.Employees sign a nondisclosure agreement and act as contractors, that is, do not advertise the link with Amazon;
- One of the employees said that the work is often routine. For example, he tracked in the records of mentioning “Taylor Swift” and arranged annotations that meant a popular singer. Teams have internal chats, where they exchange files and funny pieces of audio, for example, how someone sings in the shower;
- Two members of the team said they heard sounds similar to sexual assault. When,if they find frightening or traumatic records, they can seek psychological help.. But in fact, employees have noted that Amazon is trying not to interfere in this;
- Sometimes employees hear personal information — bank card numbers, addresses, or mobile phones. In this case, they mark the record as having “critical data” so that it is marked in the system. Sometimes they also get audio made by chance, that is, without a user command to Alex – they still have to decrypt it;
- Amazon does not give customers personal information to employees, but they see account numbers, user names, and device serial numbers. As noted by the publication, in the case of, for example, with Google assistant, employees receive audio without any personal data and with a partially distorted sound;
- Amazon said they were listening to an “extremely small” number of records, as well as “tightly monitoring” security so that employees could not identify users or use personal data. “This information helps learning speech recognition systems and understanding natural human language.”.