DeepEar listening neural network could improve virtual assistants and search engines
Fri 21 Aug 2015
They say that one of the best ways to learn is to listen. Now, there’s a program which can run in the background of a smartphone to do just that, monitoring surrounding sounds in order to ‘understand’ them better. This could inform the development of such things as virtual assistants and search engines, making them more effective at providing information specified tailored to the kinds of things people want to know.
DeepEar [PDF], developed by Bell Labs’ Nic Lane and others in Murray Hill, New Jersey, is an example of a neural network – a computer model that aims to simulate the human brain’s complexity.
Whereas most commercial examples of neural networks use the internet in order to connect to powerful computers which process the information, DeepEar just uses the smartphone’s own processors.
Not only does this conserve battery life (only using around 6% of a smartphone battery for an entire day’s ‘listening’), but it’s also potentially more secure, since the user’s personal data is not uploaded to the cloud. Obviously, the security of the individual device would still be important.
While there have been previous listening neural networks, their algorithms have struggled to make sense of noisy environments.
By contrast, as New Scientist reports, “DeepEar works by training a neural network to listen for and recognise different kinds of aural scenes, identify human speakers, recognise emotion and detect stress.”
Microsoft Research’s Dimitrios Lymberopolous in Washington State said: “This means a very responsive personal assistant that can understand both you and your environment and respond to you immediately,” adding, “This experience could make virtual assistants way more useful.”
While neural networks are not yet fully refined (something that DeepEar seeks to build on), they still have usefulness. However, sometimes their limitations can be amusing, such as some of the hilarious out-of-context mistranslations that Google Translate is capable of.
