Amazon uses audio fingerprints to prevent mentions of ‘Alexa’ in broadcast media from triggering devices

Amazon has published some interesting information about how they’re preventing Alexa devices from activating to mentions of their wake word in movies, TV shows, ads, radio, and more. An audio fingerprinting system is used to identify and store individual media mentions of Alexa that can be used to determine when the wake word should be ignored. This is done both in the cloud, to quietly turn Alexa devices back off after a media trigger, and locally on Alexa devices themselves, to completely prevent devices from waking up in the first place.

Every time an Alexa device hears its wake word, that mention is compared to digital fingerprints of known media mentions of Alexa. A small subset of media mentions, such as the upcoming Alexa Super Bowl ad, is stored on the Alexa device itself. If a wake word mention matches the subset of locally stored fingerprints, the Alexa device never even reacts to the wake word. Alexa devices cannot check against all media fingerprints locally, due to device CPU limits, so the rest are handled in the cloud.

If a known media mention of Alexa is heard that isn’t among the locally stored subset, the Alexa device will react to the wake word, but then silently turn back off once Amazon’s cloud servers identify it as a match to a known media mention fingerprint. Where things get really interesting is how Amazon is building their database of known media mentions.

When Amazon’s cloud servers receive an Alexa mention, that wake word audio is compared to a fraction of other wake words that came in around the same time. If the wake word audio matches the request of at least two other customers, then it’s identified as a media mention and used to grow the database of audio fingerprints to ignore.

From the sound of things, it seems like a live broadcast or first airing of media that mentions Alexa is much more likely to trigger your Alexa devices. However, watching reruns, on-demand, or time-shifted content that includes Alexa mentions is more likely to be in Amazon’s audio fingerprint collection and be correctly ignored by your Alexa devices. Amazon’s article about this system makes no mention of the alternate wake words, Computer, Echo, and Amazon, so it seems like using one of those will be more likely to trigger an Alexa device while playing media.

ShareTweetShare+1

Leave a Reply

Your email address will not be published.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