There probably hasn't been a time that you've consciously been thankful for deep neural networks and machine learning, but that's what Amazon is using to ensure that Alexa doesn't respond to the "Alexa" keyword when you don't want it to. 

To demonstrate this, Amazon used a real-world occurrence of when many Alexa devices would be triggered - using an advert during the Super Bowl.

There have been a number of occasions where TV adverts have sent Alexa crazy, with Echo devices in people's homes firing up and responding. This has been done for fun through a number of TV shows and it's something that happens on speaker calls, in passing conversation and many other situations.

In the case of the Super Bowl ad, however, Alexa didn't respond to the advert. Amazon said this is thanks to "acoustic fingerprinting", meaning that the Echo recognises this as an advert and not the command of whoever is in the room.

 

Sure, it's easy to schedule Alexa to ignore something predictable, but Amazon's cloud service can also pick this out on the fly. Because the audio is streamed through Amazon in the cloud, an algorithm can detect that the same audio stream has appeared from numerous devices simultaneously and can then prevent devices responding in error.

It's not perfect, however, with Amazon admitting that 80-90 per cent of devices won't respond in these circumstances as these acoustic fingerprints are created on the fly, so some are likely to be triggered.

It's clever stuff and while on a practical level it might mean that a gameshow host can't remotely instruct your Alexa to buy a new TV from your Amazon account, on a geeky level, it's great have something tangible in the machine learning bucket that feels immediately relevant.