A computer that mimics the way the brain works, and contains both optical and electronic parts, can recognize simple speech three times faster than earlier devices that used only optical components.

Reservoir computers use neural networks made of interconnected units that relay signals in recurrent, closed loops, allowing them to store information from the past and 'learn' from it. These devices are usually light-based, but Laurent Larger and his colleagues at the University of Burgundy Franche-Comté in Besançon, France, designed a simpler architecture that incorporates off-the-shelf electronics and encodes information differently using the light signals, allowing for faster performance. In speech-recognition tests, the team's device processed 1 million words per second with error rates of less than 9%.

The device could be miniaturized and put on a chip, the authors say.

Phys. Rev. X 7, 011015 (2017)