Thousands of workers employed to listen to Alexa recordings
Image credit: Dreamstime
According to an investigation by Bloomberg, Amazon has thousands of workers listening to recordings captured by customers’ smart speakers.
Amazon offers a popular range of smart speakers (Amazon Echo), which are equipped with the virtual assistant, Alexa. These speakers enter ‘listening mode’ when prompted with a wake word (usually Alexa). Once in listening mode, the speakers capture and process spoken instructions, such as to add items to a shopping list or to interact with other smart home devices.
Alexa is based on machine-learning algorithms and requires large datasets to improve its performance.
While consumer protection groups have warned of the privacy implications of the introduction of cameras and microphones in the home manufactured by the world’s largest tech companies, such as Facebook Portal, Amazon and other companies have frequently downplayed these concerns.
According to an investigation conducted and published by Bloomberg, Amazon has been improving Alexa’s natural language processing by having people listen to recordings of voice requests captured from unknowing customers addressing their smart speakers. This helps refine Alexa’s ability to ‘comprehend’ natural language in all its complexity, such as by recognising words spoken with a strong regional accent.
Data annotation can be combined with autonomous machine learning in a process known as semi-supervised learning.
Bloomberg found that thousands of employees and contractors based all around the world are engaged in data annotation. This requires them to listen to as many as 1000 voice recordings per shift, transcribe them, add notes, and feed the information back into Alexa’s machine learning algorithms. One Boston-based worker told Bloomberg that the work was mundane; specific tasks included listening to recordings of people saying ‘Taylor Swift’ and confirming that the speaker was referring to the popular singer.
The workers said that they used an internal chat room to share difficult-to-parse files or amusing recordings with their co-workers. They said that sometimes they heard instances of recordings which were embarrassing – such as singing in the shower – or deeply concerning, such as possible sexual assaults and other criminal activity. Other workers told Bloomberg that they heard as many as 100 recordings every day in which Alexa did not appear to have been deliberately activated.
The thousands of people working on the program were required to sign non-disclosure agreements which prevented them from speaking publicly about the nature of their work.
Amazon told Bloomberg: “We only annotate an extremely small sample of Alexa voice recordings in order [to] improve the customer experience. For example, this information helps us train our speech recognition and natural language understanding systems, so Alexa can better understand your requests, and ensure the service works well for everyone.”
The company said that it has a zero tolerance policy in place for abuse of the system, and that workers are not given access to the identity of people captured in Amazon Echo recordings.
Sign up to the E&T News e-mail to get great stories like this delivered direct to your inbox every day.