Is Alexa Listening?

A common concern i’ve heard about Amazon Echo devices surrounds the phrase, “Always Listening”. People aren’t only worried that the device can hear them but about who or what might be snooping in on the conversation.

According to Amazon the device is always listening out for the “wake word” which in this case is “Alexa” but, they claim that the only thing recorded and sent back to Amazon’s servers is what you specifically say to the device.

The first consideration to take into account would be where all this audio would be kept. As mentioned in my last blog the Amazon Echo Dot has quite a small amount of onboard storage (4gb of it) meaning that it wouldn’t be able to keep a long record. On top of this the quality of the storage is fairly low meaning that it is prone to wearing quickly with extensive writing.

The alternative to storing this audio on the device is that it simply uploads all the audio to the cloud and this is relatively simple to test.

A paper in 2018 by Ford and Palmer quite conclusively demonstrated that the device typically only shows great spikes in network traffic when the wake word and a command is spoken out loud to an Alexa Device during normal operation. This revelation somewhat dampens the thought that this device is constantly listening to every conversation in the room it’s placed inside. In simple terms, the device is not constantly recording everything you say.

However, to verify this I decided to conduct my own experiment. I placed an Amazon Echo Dot inside a sound proof room and played various 10 minute audio clips including;

  • A list of commands
  • Cafe noise
  • Cafe noise and a list of commands
  • Silence (control)

Each test was repeated three times leading to about 2 hours of recording to ensure that the results were reliable.

Echo Dot in a soundproof room

So, once the experiments were completed I looked at the results and they pretty much verified the conclusions of the 2018 paper.

Nine Commands Recognised by the Echo Device
Cafe Sounds being played and no large spikes
Eight Commands recognised by the Echo Device
Almost identical to the cafe noise on it’s own!

So, you may have noticed if you are eagle eyed that while I’ve noted that there were nine commands recognised in the first experiment there are eleven clear spikes. What happened there can be explained by a transcript of the audio between two hundred and three hundred seconds.

Audio SourceTranscript
Speaker Alexa, how do you make chocolate chip cookies?
Alexa Ok, How about chocolate chip cookies from Recipedia? This recipe takes ten minutes to prep and twenty minutes to cook, it serves thirty and is medium difficulty, you can hear the details or hear another recipe, what would you like to do?
Speaker[Silence, No response]
AlexaWould you like to hear the details, or another recipe?
speaker[Silence, No Response]
Alexa[Error Sound]

The Echo appears to have asked a question creating a spike and then after not getting any response back asked again, creating the second spike.

So, this seems to pretty conclusively demonstrate that the Amazon Echo doesn’t in fact send all audio back to Amazon only the phrases that are directly spoken to the device. in fact I also tested the microphone off switch and it seems to work pretty well.

However, this doesn’t get Amazon off the hook as it’s clear that audio clips that are sent back to Amazon will be used by Amazon either to improve the service or learn about you. This can involve Amazon employees listening to *some* failed audio clips to figure out what was said. I also think Amazon could probably do better work to anonymise these clips in the future.


References:

Leave a Reply

Your email address will not be published. Required fields are marked *