Those who have a speaker connected at home know that the intelligent assistants of these devices are triggered regularly by accident, while talking or watching a movie on TV.
Indeed, these software sometimes have trouble recognizing their keywords, such as “Okay Google” or “Say Siri”. The problem is that each activation causes a recording of the sound environment, and that this recording can be found on the servers of an Internet giant.
Researchers from Northeastern University and Imperial College London wanted to learn more about these misunderstandings of intelligent assistants. They bought seven smart speakers, namely a Google Home Mini 1st generation, an Apple HomePod, a Harman Kardon Invoke by Microsoft, two Amazon Echo Dot 2nd generation and two Amazon Echo Dot 3rd generation.
The devices were watered from Netflix series
These devices were placed in a box in which Netflix series soundtracks were continuously broadcast. They were continuously filmed to find out when their assistants woke up. Their network flows were also monitored to see if data was transmitted to the different clouds.
The first results of these experiments confirm the unreliability of these assistants. Not a day has gone by without them being triggered at least once by accident. The number of errors per day varies between 1.5 and 19. The most common mistakes are the HomePod and the Harman Kardon.
The longest activations were seen for the 2nd generation Echo Dot and Harman Kardon (20 to 43 seconds). In addition, more than half of the accidental activations of the HomePod and Echo Dot last more than six seconds. The risk that these devices record sensitive sound elements is therefore high.
Unsurprisingly, these erratic triggers occur when the words in Netflix dialogs closely resemble key words: “Hey Sorry” instead of “Hey Siri” (in English), “I can work” instead of “Okay Google ”,“ Colorado ”instead of“ Cortana ”,“ Kevin's car ”instead of“ Alexa ”. In short, the voice recognition systems of these devices are still far from being really efficient.
This study is not finished. The researchers will soon complete their report with other results. In particular, we will know how many of these accidental records have been sent to the cloud. The case is to follow.