German researchers have proposed a method and created hardware that
allowed them to identify more than a thousand words, in response to which
activated smart speakers from different manufacturers. The authors published
the volumetric description of work with examples on their website, and in the near future
time also plan to publish the article and assembled dataset of words
Smart columns nearly all manufacturers use a similar model of data security. They always listen to it around and trying to recognize in it the sentence active: “Alice”, “OK, Google,” “Hey Siri” or the other. This part of the work takes place completely locally, so when the perfect work smart column is passed to the company’s servers to recognize only the commands of the owner. However, the algorithms do not work perfectly and often mistaken for the phrase activate a different word that the user, for example, said in a confidential conversation by phone. The result is a messaging app sends the server a segment of speech from this word, which can be quite long. However, that part of the records analyzed software companies or even third-party contractors, these errors cause quite serious concern among experts, because that confidential conversations the user can hear the party people.
Researchers under the leadership of Thorsten Holz (Thorsten Holz) from the Ruhr University Bochum developed a method that allows to almost automatically check voice assistants on such errors and make a list of words that their algorithms interpretiruya wrong. They have collected a set of recordings of speech from movies, TV series and newscasts, totaling more than 180 hours in German only, though apart from him they have experienced speakers who speak English and Chinese. All the authors have experienced 11 devices: seven different models of smart columns on a single instance and four instances of Amazon Echo Dot because Alexa supports four different phrases activate (Alexa, Echo, Amazon, and Computer).