Hysteretic Oscillator-Based Reservoir Computing for Audio Waveform Pattern Recognition

Reservoir computing (RC), a framework for recurrent neural networks, is adept at learning the dynamics of time series data. RC, requiring less computational cost for training than traditional recurrent neural networks, is versatile in applications such as time series generation, prediction, pattern recognition, and robot control. Recently, the integration of physical system dynamics, particularly oscillatory phenomena, into RC has been explored. This study presents an RC model incorporating oscillators with hysteresis as network elements, focusing on their ubiquitous nature. In speech pattern recognition, the audio waveform, a complex vibration pattern of air, is typically preprocessed into a frequency component time series. This study, however, attempts pattern recognition by using the raw speech waveform as the direct input to the oscillator-based reservoir. The application of this model in recognizing and classifying time series vocal data is investigated, including an assessment of the oscillator elements’ bifurcation parameter on RC performance.