Birdsong Syllable Recognition Using Neuromorphic Computing

Neuromorphic computing enables highly parallel and energy-efficient computation compared to traditional CPUs. These features make it an ideal candidate for the implementation of continuous speech recognition, which requires low-latency. Audio data can be converted into sparse, spatiotemporal sequences of spikes which are input into a spiking neural network that performs speech recognition. The parallel nature of neuromorphic hardware enables a template matching network to be implemented using a large number of templates while minimizing the latency of recognition. Birdsong audio is an ideal dataset for developing speech recognition networks, as it possesses distinct syllables that can be recognized in a similar manner to human speech. Using Intel's Loihi neuromorphic chip, we are implementing a template-matching spiking neural network to perform recognition of syllables in birdsong recordings.