The Amazon Echo and other Alexa devices may soon be much more accurate in figuring out whether you’ve said something they can safely ignore, thanks to a new AI model being trained by Amazon Alexa scientist Young-Bum Kim and his colleagues. Kim’s model is a dual-layered one that screens simultaneously for relevant and irrelevant responses by sorting them into domains, much like Alexa’s current system. The difference with the current system is that it has two different models, one that detects relevant input and one that detects irrelevant and easily ignorable input. The new system boasted accuracy rates up to 6.7% better than the one currently used, in controlled testing.
The way it all works is by sorting speech into one of 21 domains, while also seeing if it pertains to any existing Alexa Skills. Speech that does not fit is checked for the possibility of the user meaning to ask for something relevant or otherwise misspeaking, and if it can’t be reconciled and identified within a category or linked to a resource, it is discarded. If this system comes to fruition, it could mean that Alexa gets better at tuning out background noise and speech, and also gets better at figuring out how best to address your request if it’s not specifically calling upon a certain Skill or aspect of the software. The team tuned the software many different ways, and for each tuning and accuracy level, ran the program back through its original training set with the new values in place 50 times in order to allow it to develop consistency.
What this boils down to is improving Alexa’s natural language processing, a cornerstone of conversational AI technology. This field of AI tech is the one that helps assistants figure out what you want even if your request is not formatted to standards, as well as figure out what sound bytes to store for training purposes and which to delete after they’ve been heard and parsed. Amazon has made great strides in this area in recent years, bringing Alexa up to par with, and in some instances even beating, the accuracy and ability of its closest competitor’s, Google Assistant and Apple’s Siri.