Metamaterials you can talk to: Speech recognition with elastic neural networks

To detect spoken commands, smart devices (for example, a speaker with Alexa or Siri) continuously convert acoustic waves to electronic signals, translate them into the digital domain, and analyze them in a signal processor. Each of these steps constantly consumes energy, imposing the need for tethered operation or large batteries. We propose to solve this problem using elastic neural networks, metamaterials consisting of arrays of coupled (potentially nonlinear) resonators. The frequencies and couplings of the resonators are optimized to maximise the speech classification accuracy (energy transmitted when excited with one word but not another). Even in purely linear metastructures, we observe binary classification accuracies exceeding 90% for a large number of pairs of words. This is demonstrated on a dataset from a large and diverse group of speakers. To attain these results, we have developed refined modelling techniques involving localised oscillations and machine learning. A unique feature of metamaterial-based speech processing is that speech classification is entirely passive, requiring no external energy. This is possible due to the very low energy dissipation of elastic waves.